Paper deep dive
Molecular LLM Agents: From Architectural Design to Scientific Autonomy
Jiatong Li, Wengyu Zhang, Weida Wang, Yuxuan Ren, Wei Liu, Chenyang Mao, Yuqiang Li, Yatao Bian, Changmeng Zheng, Xiaoyong Wei, Qing Li
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Molecular science represents an important frontier for LLM-based agents. Unlike general agents that mainly operate over natural language, code, or web environments, molecular LLM agents must perceive, reason about, and act upon chemical objects across symbolic strings, molecular graphs, 3D conformations, spectra, simulations, and wet-lab measurements. Their capabilities depend on chemically faithful molecular perception, an LLM-centered agent framework, domain-specific tool grounding, and computational or experimental feedback, in addition to planning and tool use. This work develops a conceptual framework for molecular LLM agents from two complementary perspectives. First, we introduce an architectural view of molecular-agent design, covering molecular representation and perception, the agent framework, domain-specific toolboxes, and learning and optimization. Second, we propose a scientific autonomy ladder inspired by staged autonomy in engineering systems, categorizing agents into four levels: L1 assistive or fixed workflows, L2 adaptive computational agents, L3 feedback-aware physical experiment agents, and L4 scientific-agenda agents. Together, these two perspectives establish a comprehensive framework for comparing existing molecular LLM agents, identifying missing capabilities and deployment risks, and guiding the design, evaluation, and deployment of future agents in molecular discovery workflows.
Tags
Links
- Source: https://arxiv.org/abs/2608.23104v1
- Canonical: https://arxiv.org/abs/2608.23104v1
Trouble viewing inline? Open PDF directly â
Full Text
175,790 characters extracted from source content.
Expand or collapse full text
Molecular LLM Agents: From Architectural Design to Scientific Autonomy Jiatong Li 1 The Hong Kong Polytechnic University Hong Kong, China Wengyu Zhang 2 The Hong Kong Polytechnic University Hong Kong, China Weida Wang 3 Shanghai AI Lab Shanghai, China Yuxuan Ren 4 National University of Singapore Singapore Wei Liu 5 Shanghai Jiao Tong University Shanghai, China Chenyang Mao 6 The Hong Kong Polytechnic University Hong Kong, China Yuqiang Li 7 Shanghai AI Lab Shanghai, China Yatao Bian 8 National University of Singapore Singapore Changmeng Zheng 9 * The Hong Kong Polytechnic University Hong Kong, China Xiaoyong Wei 10 * The Hong Kong Polytechnic University Hong Kong, China Qing Li 11 The Hong Kong Polytechnic University Hong Kong, China Abstract Molecular science represents an important frontier for LLM-based agents. Unlike general agents that mainly operate over natural lan- guage, code, or web environments, molecular LLM agents must per- ceive, reason about, and act upon chemical objects across symbolic strings, molecular graphs, 3D conformations, spectra, simulations, and wet-lab measurements. Their capabilities depend on chemically faithful molecular perception, an LLM-centered agent framework, domain-specific tool grounding, and computational or experimental feedback, in addition to planning and tool use. This work develops a conceptual framework for molecular LLM agents from two com- plementary perspectives. First, we introduce an architectural view of molecular-agent design, covering molecular representation and perception, the agent framework, domain-specific toolboxes, and learning and optimization. Second, we propose a scientific auton- omy ladder inspired by staged autonomy in engineering systems, categorizing agents into four levels: L1 assistive or fixed workflows, L2 adaptive computational agents, L3 feedback-aware physical ex- periment agents, and L4 scientific-agenda agents. Together, these two perspectives establish a comprehensive framework for compar- ing existing molecular LLM agents, identifying missing capabilities and deployment risks, and guiding the design, evaluation, and de- ployment of future agents in molecular discovery workflows. CCS Concepts âą Applied computingâ Bioinformatics. Keywords Molecular LLM agents, Molecular discovery, Scientific agents, Au- tonomous laboratories 1 Introduction Molecular discovery asks how to turn a desired function into a real molecular entity [65,66]. The target may involve biological activity, selectivity, toxicity, stability, or other physicochemical and functional properties [46,154]. A useful molecule must also be Scientific Goal Data Collection Agent Operation Verification Optimization 1 Help me find a molecule that ... PotencySelectivitySafetySynthesis O C1=C=C2=C1C(C)OC2 SMILES 2D Graph3D Conformer Protein Pocket Papers & Literature Databases 2 3 4 5 Results are fed back to refine design Conduct simulations or wet- lab experiments for verification SimulationsWet-lab experiments Planning Decision Memory 1) Retrieve similar ... 2) Propose scaffold ...... Agent Toolbox RetrievalDockingPropertyDFT LLM Agent Candidate 1 Candidate n ...... â Choose Lead Candidate Figure 1: End-to-end workflow of a molecular LLM agent. Starting from a scientific goal and task-specific constraints, the agent gathers multi-modal molecular evidence, plans and executes chemistry-tool operations, selects candidates, and verifies them through simulation or wet-lab experiments. Verification results are then fed back to refine the molecular design and the agentâs subsequent actions. chemically valid, compatible with multiple objectives, synthetically accessible, safe for its intended context, and supported by computa- tional or experimental evidence [11,110]. The search space is vast, evaluation is costly and uneven, and progress depends on coordi- nating the appropriate representation, model, tool, and validation signal. Author emails (numbers match the author list): 1 jiatong.li@connect.polyu.hk; 2 wengyu.zhang@connect.polyu.hk; 3 wangweida@pjlab.org.cn; 4 yuxuan.ren@nus.edu.sg; 5 captain.130@sjtu.edu.cn; 6 25047313g@connect.polyu.hk; 7 liyuqiang@pjlab.org.cn; 8 ybian@nus.edu.sg; 9 * changmeng.zheng@polyu.edu.hk; 10 * x1wei@polyu.edu.hk; 11 csqli@comp.polyu.edu.hk. * Corresponding authors: Changmeng Zheng and Xiaoyong Wei. arXiv:2608.23104v1 [cs.CL] 24 Aug 2026 PreprintMolLLMAgent Modern molecular AI has strengthened many components of this process. Graph-based models have improved structure-property prediction [158], while graph generative models support molec- ular generation and optimization in graph space [51]. Chemical language models such as ChemBERTa [20] and MoLFormer [114] learn molecular representations from SMILES [151]. MolT5 [26] and MolReGPT [67] further connect molecular structures with nat- ural language for captioning, generation, and in-context learning. Together with reaction prediction, synthesis planning, and molecu- lar optimization, these methods provide a rich stack of capabilities. Most nevertheless remain specialized models or fixed mappings, such as molecule to property, text to molecule, or candidate to score. The practical bottleneck is therefore shifting from solving an individual subtask to orchestrating subtasks into a coherent discov- ery workflow [70]. A chemist must translate a natural-language objective into molecular objects, choose among strings, graphs, conformers, and spectra, retrieve prior evidence, invoke scientific tools, reject invalid candidates, interpret noisy outputs, and decide what should happen next. Conventional molecular AI leaves most of this orchestration to the user. A model may predict, generate, or rank, while the user still decides which evidence to trust, how to repair failures, and when to continue, stop, or proceed toward synthesis and measurement. Large language model (LLM) agents offer a route from isolated models to action-oriented molecular systems. The LLM acts as a controller that interprets a goal, decomposes it into actions, in- vokes tools, observes feedback, and revises the plan. ReAct [162] made this control pattern explicit by interleaving reasoning with actions, while Toolformer [118] demonstrated language-model tool invocation. In molecular discovery, an LLM-centered system may strengthen one subtask through reasoning [71,149], retrieval or ver- ification [68], or coordinate a longer trajectory involving structure lookup, property calculation, docking, quantum chemistry, molecu- lar dynamics, retrosynthesis, and laboratory interaction. The key change is that the model participates in making decisions, selecting actions and interpreting observations rather than stopping after a single prediction or generation step. The molecular setting makes this control problem distinct from general web or software agents [99]. The agent must preserve chem- ical identity, connectivity, stereochemistry, geometry, units, and experimental conditions while moving among representations and tools. Its observations range from inexpensive heuristic scores to simulations, spectra, assays, and hardware logs, each with different uncertainty and cost. Errors can therefore propagate from a textual decision into an invalid calculation, an expensive simulation, or a physical experiment. Fluent reasoning alone is not sufficient; ac- tions and feedback must remain chemically grounded and auditable. We use molecular LLM agent to denote an LLM-centered deci- sion system that operates on molecular states, reasons over sci- entific goals, selects actions through chemistry tools or environ- ments, and uses computational or experimental observations to produce, validate, or refine molecular outcomes. Figure 1 summa- rizes this feedback-driven workflow from goal specification and data collection to agent operation, verification, and design revi- sion. ChemCrow [91] coordinates chemistry tools for structure lookup, property calculation, reaction prediction, and synthesis planning. Coscientist [10] extends the pattern toward laboratory- facing execution and feedback. These systems illustrate the move from task-specific molecular models to controllers over perception, action, tools, and evidence. Despite this progress, the field lacks a shared design framework, autonomy roadmap, and governance boundary. Existing systems differ in molecular representation, tool interfaces, memory, feed- back, optimization, and human oversight, yet are often discussed under the same agent label. This ambiguity leaves three questions unresolved. Which components should a molecular agent contain, and how should they interact? How could fixed workflows, compu- tational loops, physical experimentation, and open-ended scientific agency be distinguished? And how could increasingly consequen- tial actions be evaluated and governed? Without a common frame- work, it is difficult to compare systems, identify missing capabilities, or assess deployment risks. We address these questions through two complementary per- spectives. The architectural view decomposes molecular agents into molecular representation and perception, an LLM-centered agent framework, domain-specific toolboxes, and learning and opti- mization. It treats evaluation, safety, and trustworthy deployment as cross-cutting requirements rather than properties of the LLM alone. Section 3 separately consolidates evaluation settings for component capabilities, executable workflows, research tasks, and closed-loop discovery. The autonomy view classifies systems by the outer- most feedback loop they demonstrably close: L1 assistive or fixed workflows, L2 adaptive computational agents, L3 feedback-aware physical workflows, and L4 scientific-agenda agents. Section 4 pro- vides the operational definitions and boundary cases. Our contributions are as follows: âą We conceptualize molecular LLM agents as action-oriented systems that connect molecular state, agent control, scientific tools, and feedback. âą We develop a dual-perspective framework that links archi- tectural design to an evidence-based L1 to L4 scientific- autonomy ladder. âąWe use the framework to compare existing systems, identify capability and evaluation gaps, and characterize deployment and governance risks. 2 Architecture Design of Molecular Agents We organize a molecular LLM agent into four interacting compo- nents, as shown in Figure 2. The perception layer represents molec- ular objects and routes them among strings, graphs, geometries, images, and structured records. The agent framework reasons over scientific goals, plans actions, maintains memory, reflects on feed- back, and may coordinate specialized agents. The toolbox grounds these decisions in databases, cheminformatics, simulation, synthe- sis, and experimental interfaces. Finally, reflection, learning, and optimization convert computational or physical feedback into re- visions of molecular candidates, plans, or reusable policies. These components form a closed chain from molecular state to decision, action, observation, and optimization. PreprintConferenceâ17, July 2017, Washington, DC, USA Architectural Taxonomy Scientific-Autonomy Taxonomy Molecular representation & perception Strings, graphs, 3D conformers, spectra, and measurements become tool-readable molecular state. LLM-centered agent framework Reason, plan, remember, and reflect; coordinate specialist agents and human approval. Domain-specific molecular toolboxes Databases, cheminformatics, simulation, synthesis planning, and laboratory tools ground decisions. Learning & optimization Computational or physical evidence revises candidates, plans, tool policies, and stopping decisions. Multi-agent collaboration Specialist agents share memory, delegate tools and experiments, verify evidence, and communicate with human gates. L4 Scientific-agenda agents Cross-campaign evidence revises questions, priorities, resources, and research strategy. L3 Feedback-aware physical experiment agents Measurements change the protocol or next physical experiment without mandatory human intervention. L2 Adaptive computational agents Digital evidence changes the next candidate, tool, plan, or stopping decision. L1 Assistive / fixed workflows Human defines the goal and path; the system executes predefined steps. Safety & trustworthy deployment Constraints · permissions · fail-safe execution Persistent provenance & memory State · evidence · versions · audit trail Evaluation & calibration Task success · uncertainty · cost · reproducibility Human oversight & approval Objectives · gates · preferences · final judgment Figure 2: Two complementary taxonomies for molecular LLM agents. The architectural taxonomy (left) organizes agent systems around molecular representation and perception, an LLM-centered agent framework, domain-specific molecular toolboxes, and learning and optimization, together with multi-agent collaboration and cross-cutting deployment requirements. The scientific-autonomy taxonomy (right) classifies systems by the outermost feedback loop they can reliably close without mandatory human intervention, ranging from L1 assistive or fixed workflows to L4 scientific-agenda agents. 2.1 Molecular Representation and Perception Perception is the interface through which a molecular agent turns a chemical object into a state that an LLM-centered controller can read, edit, verify, and pass to tools. This interface is more constrained than ordinary text perception: a molecule has atom identities, bond orders, aromaticity, charge, stereochemistry, con- formers, electronic effects, and task-dependent physical context. Consequently, the agent does not perceive âthe molecule itselfâ, but a representation of it. The choice of representation influences which chemical facts are explicit, which facts must be inferred, and which actions are comparatively easy or brittle; these effects also depend on the model, task, and validation tools surrounding the representation. 2.1.1Representation Substrates: Strings, Graphs, Geometry, Images, and Structured Text. The most common entry point is a 1D chemical line notation. SMILES serializes a molecular graph into an ASCII string and remains attractive because it is compact, parser-friendly, and directly compatible with sequence models [151]. InChI offers another standardized identifier oriented toward chemical databases and interoperability [42]. For LLM agents, these strings are con- venient action interfaces: a controller can generate a candidate SMILES, call a toolkit to parse it, compute descriptors, search a database, or send it to a retrosynthesis service. Their weakness is that the graph is implicit in a traversal. Branches, ring closures, aromaticity, and stereochemical marks must be reconstructed from the sequence, and different valid strings can describe the same mol- ecule. This is acceptable for fast screening and tool calls, but fragile for tasks that require exact topology editing [72] or long-range structural reasoning. SELFIES changes the failure mode by defining a robust string representation in which every SELFIES string maps to a chemi- cally valid molecule under the representationâs constraints [58]. This makes SELFIES especially useful when an agent repeatedly samples, mutates, or optimizes molecular strings, because local gen- eration errors are less likely to collapse into unparsable outputs. The guarantee, however, is a validity guarantee rather than a guarantee of stability, synthesizability, or task quality. Fragment-level variants move the perceptual unit closer to how chemists design molecules. Group SELFIES extends SELFIES with group tokens for common functional groups or substructures [18], while SAFE represents a molecule as an unordered sequence of connected fragment blocks that remains compatible with SMILES parsers [100]. MolLingo and mCLM continue this direction by using molecule-native or synthesis-friendly building blocks for LLM-powered agents and chemical language models [25,98]. These representations are less character-centric and more action-centric: the agent can operate PreprintMolLLMAgent on scaffolds, linkers, substituents, motifs, or functional modules instead of only on individual symbols. Two-dimensional molecular graphs expose the topology directly. Atoms become nodes, bonds become edges, and node or edge fea- tures encode element type, valence, charge, aromaticity, bond order, and stereochemistry. This makes graph encoders natural for prop- erty prediction and structure-aware generation; directed message passing is a representative example of this family [158]. The diffi- culty is that a graph is not natively an autoregressive text sequence. An LLM agent therefore needs either a graph encoder whose con- tinuous output is projected into the language model, or a discrete graph-to-text interface that turns nodes and bonds into tokens. Re- cent structured molecular languages pursue the latter route. MolJ- SON encodes atoms and bonds in a JSON schema designed for LLM reasoning [117], and MoleCode uses explicit node-edge-subgraph primitives so that connectivity and stereochemical information are visible in the context window [157]. The broader lesson from these works is that structure-sensitive reasoning improves when the relational structure is not hidden inside a linearized string. Three-dimensional representations add geometry. Coordinates, distance matrices, conformer ensembles, protein pockets, docking poses, or crystal structures are necessary when the relevant prop- erty depends on spatial arrangement rather than only connectivity. SchNet showed how neural models can learn from interatomic distances for quantum interactions [119], and GraphMVP aligns 2D graph representations with 3D geometry to transfer conforma- tional information into molecular graph encoders [83]. For agents, 3D perception is expensive but often decisive: docking, molecular dynamics, transition-state reasoning, binding-pose inspection, and materials simulation require coordinates and physical units that a line notation does not provide. The practical problem is that a molecule may have many low-energy conformers, and downstream decisions can be sensitive to how those conformers are generated, ranked, and passed between tools. Images provide a fourth perceptual channel [73]. A rendered skeletal formula is close to what chemists see in papers, patents, and lab notebooks. MolScribe treats molecular structure recog- nition as image-to-graph generation [112], and MolSight studies progressive visual pretraining over molecular diagrams for prop- erty prediction [7]. Image perception is valuable when the agent must read literature or multimodal records, but it also introduces optical recognition errors: atom labels, wedge bonds, charges, and crowded ring systems must be converted back into a valid graph before most tools can use them. Experimental observations form a further perceptual substrate once an agent interacts with instruments or laboratory records. Spectra, chromatograms, assay tables, images, time series, protocol events, and robot logs describe a molecular system under particular conditions rather than molecular identity alone. They therefore require the agent to preserve units, acquisition settings, sample identifiers, uncertainty, and provenance together with the mea- sured values. Coscientist interprets UVâVis measurements inside a physical task [10]; LLM-RDF uses reaction yields and spectral analysis in an end-to-end synthesis-development workflow [116]; and ORGANA combines visual feedback with experiment execu- tion and reporting [22]. These systems illustrate why experimental records should be represented as state-bearing observations, not flattened into unqualified text. 2.1.2Tokenization: The Agentâs Molecular Action Interface. Once a molecule is written as text or structured text, the LLM still sees only token IDs. For LLM-based molecular agents, tokenization is not a neutral preprocessing step: it defines the granularity at which the agent can act on molecules, the way errors propagate through tool chains, and whether the agent can later read and reflect on its own molecular edits. The question is distinct from tokenization for spe- cialized chemical language models such as ChemBERTa [20], MoL- Former [114], or SELFormer [169], which train dedicated encoders on domain-specific corpora and can choose tokenizers optimized for their own architecture. An LLM agent, by contrast, typically operates through a general-purpose language model whose tok- enizer is fixed at pretraining. Molecular strings must pass through this given tokenizer and remain actionable on the other side. The agent-level consequences of this constraint vary across autonomy levels. Character-level molecular tokens and the L1 baseline. A generic subword tokenizer (BPE, WordPiece, Unigram) trained on natural- language corpora can split multi-character elements such asCland Br, scatter bracketed atoms and stereochemical marks across several tokens, and merge chemical punctuation with atoms in arbitrary ways [135]. At L1, where a human or fixed workflow closes each consequential transition, this can be tolerable: the agent generates a SMILES candidate, and a human review or deterministic valida- tion catches parsing errors before they reach a downstream tool. Atom-wise tokenizers that preserve atom boundaries and brack- eted atoms improve this baseline, and Smirk decomposes bracket atoms into glyph-level tokens to retain open-vocabulary coverage for organometallics, coordination complexes, and other chemistry that general tokenizers replace with undifferentiated unknown to- kens [147]. Even at L1, coverage matters because agents may query patents, catalysts, materials, metal complexes, polymers, salts, and tool-produced intermediates whose chemistry falls outside the clean organic subset seen during pretraining. Fragment-level tokens and L2 iteration efficiency. When the agent moves to L2 and autonomously iterates through cycles of proposal, evaluation, and revision, tokenization begins to shape the optimiza- tion dynamics. A character-level action edits one atom symbol; re- placing a substituent on a scaffold may require five to eight sequen- tial edits. Fragment-level tokenizers, including Group SELFIES [18], SAFE [100], and mCLM [25], elevate the action unit to a chemically coherent motif such as a functional group, scaffold, or linker. The same edit can then be expressed in a single step. This can shorten the action sequence needed to express a scaffold-level edit and may simplify search or credit assignment when the model and objective are aligned with the chosen fragments. The realized gain in plan- ning or computational efficiency is system-dependent, however, and should not be attributed to token granularity alone. Fragment-level tokenization also changes the error profile. At the character level, a mismatched ring-closure digit or unbalanced bracket can cause an RDKit [8] parsing failure. Descriptor computation and docking preparation then cannot proceed, so the workflow halts. The agent must detect the failure, diagnose its source, and retry, consuming PreprintConferenceâ17, July 2017, Washington, DC, USA Table 1: Common molecular and experimental perceptual substrates for LLM-based molecular agents. SubstrateTypical formWhat it makes explicitAgent-level use and main risk Line notationSMILES, InChICompact identity and parser-compatible connectivityFast generation, search, and tool calls; topology and stere- ochemistry must be recovered from a brittle sequence. Robust or fragment stringSELFIES, Group SELFIES, SAFE, molecule-native frag- ments Validity constraints or chemically meaningful blocks Useful for repeated editing and optimization; validity does not ensure synthesizability or property quality. Graph or structured textMolecular graph, MolJSON, MoleCodeAtoms, bonds, local topology, explicit node and edge iden- tities Better for exact edits and graph reasoning; requires graph encoders or verbose structured contexts. 3D geometryConformers, coordinates, pockets, docking poses Distances, angles, chirality, spatial contacts, physical stateNecessary for binding, quantum, and simulation tasks; expensive and conformer-sensitive. ImageSkeletal diagram or scanned chemical figureHuman-readable structural drawing and visual stereo- chemistry cues Connects agents to papers and notebooks; must be recog- nized into graph or string form before most tools can act. Experimental recordSpectra, chromatograms, assays, time series, protocol events, robot logs Measured response, conditions, units, acquisition context, and execution state Grounds interpretation and laboratory control; sample mismatch, lost metadata, or overconfident peak assign- ment can corrupt later decisions. planning budget. SELFIESâs validity guarantee [58] and fragment- level constraints reduce the space of syntactically malformed out- puts. They can therefore limit one source of error propagation at the perception-to-tool interface, while leaving stability, synthesis feasibility, and task quality unresolved. This benefit may be missed by language-modeling evaluation but can matter for L2 agents that generate and evaluate many candidates in a campaign. Tool aug- mentation can otherwise degrade chemistry problem solving when inputs are malformed [78, 166]. Structure-discretized tokens and the transparency-efficiency trade- off. A newer family of tokenizers discretizes molecular structure rather than strings. UniMoT uses a molecule encoder, a causal Q- Former, and vector quantization to map molecular graph features into discrete tokens that can be added to an LLM vocabulary [39]. AtomDisc and VQ-Atom similarly use graph or geometry context to quantize atom-level chemical neighborhoods [56,175]. For agents, this can reduce the interface gap between continuous molecular en- coders and symbolic action traces and compress complex substruc- tural edits into single token predictions. The cost is reflective opac- ity. Agents reflect on their action trajectories by reviewing previous tool calls, diagnosing failures, and deciding what to retry [75,133]. This reflection operates in the LLMâs text space: character-level and fragment-level tokens are human-readable, so the agent and a human auditor can inspect which molecular edit was attempted and what went wrong. Structure-discretized tokens are opaque. An agent may record âI predicted code 437âcode 892â without being able to express the corresponding structural change unless it invokes an external decoder. MolLingo [98] proposes molecule- native representations that retain structural meaning, while La- tentChem [164] replaces verbose textual chain-of-thought with latent thinking; both illustrate the design tension. At L2, where the agent must self-correct during iterative optimization, reflective opacity can be a significant bottleneck. 2.1.3Perception as a Routing and Interaction Layer. Current molec- ular agents rarely rely on a single representation throughout a task. They route among representations according to the action being taken: generating a SMILES candidate, parsing it into a graph for editing, converting to 3D for docking, converting back to SMILES for a database query, and rendering an image for human review. Chem- Crow couples an LLM controller to chemistry tools so that natural- language requests can be translated into such operations [91], and ChemAgent adds memory and tool-use policies so that the per- ceived state includes prior tool outputs, evidence, and task his- tory [133]. Routing begins with entity resolution. A user-supplied name, CAS number, or drawn structure is mapped to a canonical molecular representation through name-to-structure conversion with resources such as OPSIN [89] or PubChem [55], followed by protonation-state assignment and tautomer selection. Routing continues through every later representation transition. At L1, a human or fixed workflow directs each transition and catches errors before they propagate; the routing layer is effectively a bidirec- tional format translator. As autonomy rises, routing must become an autonomous state-management system with two capabilities that are largely unaddressed in existing architectures: conversion fidelity and bidirectional tool-facing perception. Conversion validation. Each representation transition can lose information: stereochemistry may not survive a SMILES-to-graph- to-SMILES round trip; a conformer generator may fail silently; a structured-text encoding may truncate features that exceed the context-window budget. Without validation checkpoints at each conversion boundary, the agent may silently operate on a different molecule from the one it started with. MolJSON [117] and Mole- Code [157] make relational structure visible in the context window, which supports conversion-aware perception. Most agent archi- tectures still lack systematic checks that atom count, bond count, stereochemistry, and charge survive each transition. At L2, where the agent controls the routing, these checkpoints become manda- tory infrastructure; a single undetected loss in cycle 3 of a 10-cycle optimization corrupts all subsequent cycles. At L3, conversion vali- dation extends to physical signals: the perception layer must verify both structural preservation and correct parsing, compound assign- ment, and condition annotation for each experimental readout. Output perception and feedback perception. Before invoking a tool, the agent must transform its internal state into the toolâs ex- pected input format: a 3D SDF file for docking, canonical SMILES for a database query, charge and spin multiplicity for quantum chemistry [182], or a valid graph with explicit hydrogens for ret- rosynthesis. Many agent failures originate at this output-perception boundary: the LLM reasons correctly about which tool to call but constructs an invalid input. ChemHAS [78] and TRACE [75] address this through self-correction at the tool interface, but recognizing these as perception failures rather than reasoning failures enables more targeted evaluation. In the reverse direction, a tool output PreprintMolLLMAgent such as a docking score, ADMET prediction, DFT energy, or spec- tral match must return to the agentâs reasoning space with context. The agent must determine whether the output corresponds to the intended molecule, quantify its uncertainty, and identify conflicts with other signals. Tooling-or-Not-Tooling [166] shows that tool augmentation can help or hurt depending on context. The percep- tion layer should therefore assess the reliability of each feedback signal before incorporating it into the agent state. At L3, feed- back perception faces a distinct challenge: experimental signals are inherently ambiguous, and the perception layer must preserve this ambiguity rather than collapsing to a single interpretation, since premature disambiguation can produce false experimental conclusions. At L4, feedback from different campaigns, tools, and scoring functions must be compared on a common perceptual ba- sis, which requires unified state representations and provenance tracking across projects. Thus, molecular perception should route among multiple repre- sentations, remain aware of tokenizer effects, validate conversions explicitly, and support bidirectional interaction with tools. The downstream controller can plan reliably only when its perceptual state preserves the chemical constraints needed by the next action. This requirement scales from human-checked format translation at L1 to cross-campaign state unification at L4. 2.2 Agent Framework The agent framework is the LLM-centered controller between molecular representation and scientific action. It selects tools or subagents, interprets their outputs, and decides whether a trajectory should continue, revise, stop, or request approval. ChemCrow [91] and CACTUS [93] illustrate chemistry-tool control; MDCrow [13] and LLaMP [19] extend it to longer computational workflows; and Coscientist [10], LLM-RDF [116], and Tippy [27, 28] move toward physical or laboratory-facing decisions. 2.2.1 Overview: The Molecular Controller. A molecular controller is judged by the scientific objects and actions it can produce, not by fluent text alone. Its outputs must remain chemically valid, its decisions must be grounded in noisy or costly evidence, and its internal state must preserve the molecular representation needed by downstream tools. Most systems implement an observe, plan, act, and revise loop through five faculties: reasoning, planning, memory, reflection and self-correction, and multi-agent collaboration. Figure 3 cross-tabulates papers by these faculties and by the highest demonstrated autonomy level under Section 4. Each paper receives one global level that is reused across faculties: the heatmap counts it once per relevant faculty, whereas the yearly bars count it once overall. 2.2.2 Reasoning. Reasoning turns a molecular goal into hypothe- ses, intermediate conclusions, and candidate actions. Early chem- istry agents often express this reasoning as natural-language chain-of-thought over names, SMILES strings, reaction descrip- tions, or retrieved text. Such reasoning is interpretable, but it is fragile: a small mistake in valence, stereochemistry, units, reaction feasibility, or tool-input syntax can invalidate a fluent explana- tion. Chemistry-oriented LLMs therefore strengthen the reasoning substrate through domain training, memory, and molecular lan- guage modeling, as illustrated by ChemAgent [133], ether0 [95], mCLM [25], and comprehensive molecular design language mod- els [168]. One direction replaces linear reasoning with search-augmented reasoning. Molecular design is combinatorial, and many tasks re- quire comparing multiple hypotheses rather than committing to one generated chain. Monte Carlo Thought Search [127] explores catalyst-design reasoning paths with tree search, while ChemRea- soner [128] searches an LLMâs chemistry knowledge space and grounds the search with quantum chemical rewards. CheMatA- gent [152] learns chemistry and materials tool-use policies through tree search based training, DrugMCTS [161] combines retrieval, multi-agent roles, and Monte Carlo Tree Search for drug repurpos- ing, and Agents-on-a-Tree [173] coordinates pathwise molecular optimization. A molecular reasoning branch can therefore be eval- uated by docking, property prediction, adsorption energy, reaction barriers, quantum calculations, or other physical feedback instead of linguistic plausibility alone. A second direction makes reasoning structured and executable. Instead of producing only plain text, the agent emits tool calls, action sequences, code, protocol steps, or parameterized workflows that can be parsed, executed, checked, or replayed. ChemAc- tor [177] converts unstructured synthesis procedures into machine- executable chemical action sequences, MT-Mol [54] decomposes molecular optimization into tool-based reasoning, and DrugPi- lot [74] uses parameterized reasoning over multimodal drug- discovery information. Coscientist [10], El Agente [182], DrugA- gent [82], and Chemist-X [16] expose reasoning through executable code, workflow traces, or computer-aided design interfaces. A third direction asks whether the agent should reason in text at all. Text is inspectable, but it is lossy for graphs, conformers, stereochemical relations, and 3D interactions. MolLingo proposes molecule-native representations for LLM scientific agents [98], and LatentChem replaces verbose textual chain-of-thought with latent thinking and a dynamic perception loop [164]. This creates a design trade-off: text supports human inspection, whereas molecule-native or latent states may preserve structural information that is hard to express in prose. The boundary between reasoning and tool use is also adaptive rather than fixed, since tool augmentation can help some chemistry tasks and hurt others [166]. Reasoning becomes autonomy-relevant only when its evidence changes a later scientific action. ChemNavigator [108], for example, extracts design rules within a bounded computational campaign and is therefore an L2 boundary case rather than evidence of cross-campaign L4 discovery. 2.2.3 Planning. Planning turns a reasoning outcome into an exe- cutable trajectory. For molecular agents, a plan must specify how information moves across representations, tools, constraints, and feedback signals. A goal may require literature retrieval, candidate generation, filtering, docking, retrosynthesis, simulation, protocol generation, or laboratory execution. Thus, planning lengthens the agentâs horizon from one-shot tool use to in-silico campaigns. A common pattern is task decomposition and workflow orches- tration. M 4 olGen [76] studies multi-stage generation under precise multi-property constraints, while Prompt-to-Pill [144], PharmA- gents [30], MADD [125], and FROGENT [105] organize broader PreprintConferenceâ17, July 2017, Washington, DC, USA L1 Assistive or fixed L2 Adaptive computational L3 Feedback-aware physical Reasoning Planning Memory Reflection and self-correction Multi-agent collaboration Agent-framework faculty 1181 9124 770 4100 8114 L4 Scientific agenda No papers yet evidence gap 2023202420252026 0 5 10 15 20 25 30 Number of papers 1 4 13 9 2 7 11 7 1 1 2 n=4 n=12 n=26 n=16 L1L2L3 Darker shade = more faculty-level assignments (a) Controller faculties by autonomy level (b) Papers by publication year Figure 3: Two views of the 58 surveyed papers by highest demonstrated autonomy level. The faculty heatmap counts 96 paperâfaculty assignments, while the yearly stacked bars count each paper once; hatching marks the L4 evidence gap. drug-discovery pipelines into stages or agents. These systems show that generation, scoring, synthesis analysis, and reporting work closely together: each stage changes the state the planner must pass to the next module. Another pattern is interleaved reason-act planning. Rather than drafting a complete plan once, the agent alternates between lo- cal reasoning, tool invocation, observation, and revision. Chem- Crow [91] and CACTUS [93] exemplify this mode for chemistry tasks, LLaMP [19] extends it to materials retrieval and simulations, and MDCrow [13] applies it to molecular-dynamics workflows. Longer tasks motivate hierarchy: El Agente [182], MASTER [115], robotic ChemAgents [126], and Tippy [27,28] separate high-level scientific intent from mid-level workflow control and low-level execution. Planning also includes tool selection and action-policy design. In molecular agents, choosing RDKit, docking, retrosynthesis, DFT, molecular dynamics, a database, or a verifier is already a scien- tific decision. ChemHTS [77] studies hierarchical tool stacking, and ChemHAS [78] improves chemistry-tool performance through agent stacking. TRACE [75] frames lead optimization as resource- aware planning, CheMatAgent [152] learns tool-use policies, and Mozi [14] represents governed autonomy through state-aware skill graphs. The planner should therefore be treated as a policy over admissible scientific actions, not as a free-form text generator that happens to call tools. The planning loop closes at different levels. At L2, agents propose candidates or workflows, evaluate them with predictors or simula- tors, and refine the next action, as in dZiner [4], ChatMOF [53], and MDCrow [13]. At L3, Coscientist [10] and ORGANA [22] demon- strate planning and physical execution with feedback-aware recov- ery, perception, or analysis. LLM-RDF [116] and robotic ChemA- gents [126] demonstrate the stronger iterative form of L3 because measured reaction yields or catalyst performance determine a sub- sequent physical experiment. AutoLabs [106] remains L2 under our evidence rule: it evaluates multi-agent protocol generation and self-correction, but reports a critical human verification step be- fore instrument execution. Tippy [27,28] also remains L2 because its reported evaluation does not establish autonomous physical execution. At L4, planning would become research-program de- sign, where agents select problems, allocate resources, compare campaigns, and extract reusable rules. 2.2.4 Memory. Memory gives the controller persistence beyond the current prompt. Molecular discovery may involve repeated molecule edits, failed tool calls, changing constraints, simulation parameters, synthesis attempts, assay evidence, and human deci- sions. A context window can hold part of this state, but it does not provide durable, searchable, provenance-aware scientific mem- ory. ChemAgent [133] uses self-updating memories to improve chemical reasoning, while modular drug-discovery agents [102] and TRACE [75] show why long-horizon tasks require preserving intermediate state and action history. Molecular memory is structured rather than conversational. It may store molecules, conformers, pockets, score vectors, dock- ing poses, reaction conditions, failed candidates, tool parameters, literature snippets, assay records, protocol versions, and design rationales. ChemAgent [133] is a representative typed-memory architecture with planning, execution, and knowledge memories; DrugPilot [74] maintains a parameterized memory pool for multi- modal drug-discovery information; and El Agente [182] uses mem- ory inside a hierarchical quantum-chemistry workflow. Retrieval-augmented generation is another form of external memory. Chemist-X [16] retrieves literature and database evidence for reaction-condition recommendation, ChatDrug [84] combines retrieval with domain feedback for conversational drug editing, and RAG-enhanced collaborative agents [62] use retrieval for drug- discovery reasoning. Agent-based learning from literature [3] ex- tracts structured materials data from papers, while DrugAgent [48] and DrugMCTS [161] illustrate knowledge-graph and retrieval memory for drug-target or repurposing tasks. Provenance is essen- tial: the agent should retain a fact together with its source and any later contradictory evidence. Experience memory connects agent design to optimization. Pre- vious failures can prevent repeated invalid edits, bad docking setups, PreprintMolLLMAgent weak analogs, and tool-instruction errors. TRACE [75] stores in- struction and result histories as action-level experience, Augmented Memory [38] uses experience replay for sample-efficient de novo design, ExLLM [113] uses experience-enhanced optimization, and Mozi [14] stores reusable procedures in skill graphs. Memory also introduces risks: stale literature, noisy proxy scores, contaminated examples, and incorrect tool outputs can be amplified if the con- troller does not decide what to write, retrieve, forget, and trust. These memory functions support progressively longer and more reproducible workflows, but stored context alone does not raise autonomy unless it changes a later scientific decision. 2.2.5Reflection and Self-Correction. Reflection converts feedback into correction. In general LLM agents, reflection often means tex- tual self-critique. For molecular agents, this is insufficient because a fluent critique may miss an invalid structure, infeasible reac- tion, wrong unit, malformed docking input, unstable simulation, or unsafe protocol. A molecular controller should therefore reflect against external signals such as validity checks, property predictors, docking, retrosynthesis, spectra, assays, failed tool calls, hardware feedback, and expert review. A major use is iterative molecular editing and optimization. Chat- Drug [84] combines conversational editing with retrieval and do- main feedback, AgentDrug [61] uses domain feedback to steer zero-shot molecular optimization, and Probe-Before-You-Edit [160] uses structure-based feedback before editing molecules. DrugAs- sist [163], GeLLM 3 O [23], and ExLLM [113] similarly condition later proposals on observed weaknesses of earlier candidates. In these systems, reflection is not an explanation after generation, but part of the optimization dynamics. Another pattern separates generation from verification. MT- Mol [54] uses specialized tool-based roles including verifier and reviewer functions; ChemActor [177] adds structured action se- quences and multi-round review for synthesis extraction; and Chem- Labs [156] uses multi-agent checking for multimodal chemistry rea- soning. Debate is a stronger variant: Mol-Debate uses disagreement among agents to improve molecular structural reasoning [176], and collaborative expert LLMs expose trade-offs in multi-objective optimization [167]. However, debate is useful only when agents bring diverse evidence, tools, or objectives; otherwise it can amplify shared errors. Reflection also applies to tool failures. Invalid SMILES strings, malformed arguments, missing database fields, incompatible files, unstable simulations, and contradictory scorers can all corrupt downstream planning. ChemHAS applies self-correction at the chemistry-tool interface [78], and TRACE reuses previous tool- instruction failures to refine future actions [75]. AutoLabs [106] extends self-checking to the generation of hardware-ready exper- imental procedures, but its reported pre-execution human verifi- cation gate keeps the evaluated workflow at L2 under our rubric. The autonomy contribution of reflection therefore depends on the source of the feedback and on which later decision it is allowed to revise, not on the presence of a self-critique step. 2.2.6 Multi-Agent Collaboration. Multi-agent collaboration dis- tributes controller functions across specialized roles. This design is natural because discovery spans heterogeneous artifacts and skills: literature, structures, protein pockets, reaction schemes, simula- tions, spectra, protocols, and lab readouts. A single controller may struggle to maintain all contexts, whereas specialized agents can separate literature search, generation, scoring, synthesis planning, analysis, verification, and human communication. The relevant de- sign criterion is whether the division of labor matches the molecular task, not the number of agents. Role-specialized pipelines decompose discovery into stages. Phar- mAgents [30], Prompt-to-Pill [144], MADD [125], FROGENT [105], and M 4 olGen [76] assign different agents or stages to target anal- ysis, generation, scoring, synthesis, and reporting. At L3, agents can map directly onto laboratory roles. Coscientist [10] coordi- nates specialized planning, search, code, and automation modules. LLM-RDF [116] assigns agents to literature scouting, experiment design, hardware execution, spectrum analysis, separation instruc- tion, and result interpretation, while robotic ChemAgents [126] coordinates literature, experiment design, computation, and robotic execution in an iterative materials campaign. ORGANA [22] com- bines task planning, visual feedback, robot control, and reporting. AutoLabs [106] and Tippy [27,28] define useful laboratory-facing roles but remain L2 here because their reported evaluations retain a pre-execution human gate or do not establish autonomous physical execution. Many systems use an orchestrator or supervisor. Mozi [14] main- tains governed autonomy through state-aware skill graphs. MAS- TER [115] uses hierarchical multi-agent reasoning for functional- materials discovery. El Agente [182] organizes quantum-chemistry workflows through hierarchical control. This authority structure improves traceability, since decisions can be attributed to a super- visor, specialist, tool, or human gate, and it supports permissioning for high-risk actions. Collaboration can also improve critique [178]. Mol-Debate [176], MT-Mol [54], ChemLabs [156], and MASTER [115] use multi- agent disagreement, verification, or peer review as quality control. Agents-on-a-Tree [173], DrugMCTS [161], multi-GPT-agent re- inforcement learning [45], and collaborative expert LLMs [167] coordinate agents over shared chemical search spaces. Humans can also be first-class collaborators: collaborative structure-based drug design [31] and ORGANA [22] keep experts in the loop for objectives, approvals, preferences, and final judgment. Role specialization can support any level; the classification depends on the strongest evidence-conditioned action performed by the team as a whole. 2.2.7 Cross-Cutting Synthesis. Figure 3 contains 96 faculty-level assignments: 39 at L1, 48 at L2, and 9 at L3. Fixed tool chains, prede- fined role pipelines, and review-only ensembles remain L1 unless evidence changes a later scientific decision. Feedback-aware physi- cal workflows qualify as L3, with measurement-driven experiment selection identified as iterative L3. No surveyed paper meets the L4 criterion. The five faculties are therefore enabling mechanisms rather than autonomy levels. Their role changes with feedback source and action authority: memory progresses from retrieval to experimental state, while collaboration progresses from review ensembles to divisions of laboratory work. As authority increases, permissioning, PreprintConferenceâ17, July 2017, Washington, DC, USA traceability, tool constraints, and approval gates become part of the controller itself. 2.3 Domain-Specific Molecular Toolboxes Molecular toolboxes translate language-level plans into executable chemical operations over structures, databases, simulations, spec- tra, and laboratory protocols. ChemCrow [91] and CACTUS [93] illustrate chemistry-tool control, Coscientist [10] extends it to- ward experiments, and ToolUniverse [32] treats tool composition as reusable scientific infrastructure. The toolbox is therefore part of the agent architecture: it defines both the available actions and the observations that can enter the control loop. Those observations differ in cost and reliability. Computational scores, assay records, spectra, and robot logs must be interpreted within their operating conditions. Tool access alone is not ben- eficial [166]; self-correction, experience, and workflow structure determine whether evidence improves a later decision [75,78,174]. Accordingly, fixed calls remain L1, evidence-adaptive computational tool use is L2, and autonomous physical execution with feedback is L3. The criterion is scientific authority exercised through tools, not the number of tools. 2.3.1 Molecular state construction and knowledge grounding. Entity resolution and molecular state construction. A molecular agent first has to determine which chemical object it is acting on. User inputs may name a compound, give an IUPAC name, pro- vide a CAS number, mention a target or protein family, quote a paper fragment, or describe an assay. The agent must resolve these surface forms into identifiers and representations such as SMILES, InChI, PubChem CID, ChEMBL ID, UniProt ID, PDB struc- tures, or assay records. OPSIN, PubChem, ChEMBL, RCSB PDB, UniProt, ZINC, and Materials Project belong in this grounding layer [9,49,50,55,89,142,170]. A wrong molecule, target, pro- tonation state, or assay condition can invalidate the rest of the workflow. ChemCrow [91] and CACTUS [93] rely on name-to- structure conversion and database grounding as entry points for tool use. ChemAgent [133], Chemist-X [16], and LLaMP [19] show how structured state construction can be combined with memory or retrieval. Cheminformatics validation and descriptor computation. Chemin- formatics toolkits provide many of the low-level operations that molecular agents need. RDKit [8] and Open Babel [101] parse SMILES and other formats, canonicalize structures, validate va- lence and aromaticity, preserve or check stereochemistry, com- pute descriptors, generate fingerprints, support similarity search, and convert file formats. These operations matter because LLM- generated molecular strings can be fragile, invalid, or underspec- ified. Once a text output becomes a checkable molecular object, the tool can act both as calculator and validator. ChemCrow [91] and CACTUS [93] use such operations for chemistry problem solv- ing. MT-Mol [54], ToolMol [179], CheMatAgent [152], TRACE [75], and MolClaw [174] place them inside longer optimization or tool- planning loops. Evaluation can draw on established benchmarks for molecular design and property prediction, including Molecu- leNet [154], GuacaMol [11], MOSES [110], and TDC [46]. Database and literature grounding. Molecular agents also need structured scientific memory. Databases and literature tools pro- vide evidence about identity, target annotations, protein structures, bioactivity values, assay metadata, commercial availability, reaction precedent, and experimental results. Molecular databases differ from general web search because they preserve identifiers, units, curation history, and provenance. PubChem and ChEMBL ground compounds and bioactivity data; RCSB PDB and UniProt connect agents to protein structures and protein knowledge; ZINC and Materials Project support purchasable-molecule and materials retrieval [9,49,50,55,139,142,170]. Chemist-X [16] uses retrieved evidence for reaction-condition recommendation. LLaMP [19] grounds materials reasoning in high-fidelity retrieval, and agent- based literature learning extracts structured materials datasets from papers [3]. Retrieval-augmented drug-discovery agents extend this pattern to drug-target reasoning and collaborative molecu- lar decisions [48,62]. The difficult parts are entity linking, unit normalization, conflicting evidence, and versioned provenance. 2.3.2 In-silico evaluation and simulation tools. Low-cost oracles and structure-based screening. In-silico evalua- tion tools let an agent test a hypothesis before synthesis or wet-lab work. Low-cost tools include descriptor calculators, drug-likeness filters, synthetic-accessibility scores, ADMET predictors, and toxic- ity predictors. Structure-based tools add docking, scoring, binding- pose analysis, and virtual screening. AutoDock Vina, GNINA, and DiffDock, together with benchmarks such as PDBbind and CASF, are common substrates for pose prediction, scoring, ranking, and screening [21,94,130,140,148]. These tools make iterative opti- mization practical: the agent proposes candidates, evaluates them with proxy objectives, and revises them. CACTUS [93] and MT- Mol [54] incorporate tool-based evaluation into agent workflows. Several drug-discovery agents combine generation, filtering, dock- ing, retrieval, and optimization in more specialized settings [31, 61,74,125,160,161,174,179]. These scores should still be treated as decision signals, not as ground truth. Docking scores approx- imate binding, ADMET models depend on training distributions, and heuristic filters can reject unusual but useful chemistry. Quantum chemistry and atomistic simulation. Quantum-chemical and atomistic simulation tools provide physical feedback at higher cost. They compute geometries, single-point energies, vibrational frequencies, reaction energetics, electronic structures, and mate- rial properties. ASE, xTB, Psi4, ORCA, Gaussian, pymatgen, and LAMMPS expose these calculations through programmable inter- faces, which makes them usable inside agent workflows [5,6,29, 43,96,104,107,137]. For molecular agents, these tools ground part of the reasoning in physics rather than text or heuristic scores. El Agente [182] demonstrates autonomous quantum chemistry. Chem- Reasoner [128] uses quantum-chemical feedback to guide heuristic search over an LLMâs chemistry knowledge space. ChemGraph exposes molecular simulation workflows as agent-compatible com- putational chemistry tasks [109]. Monte Carlo Thought Search [127] explores reasoning paths in catalyst design, and CheMatAgent [152] learns tool-use policies for chemistry and materials science. These systems reveal a demanding control problem: the agent must choose method, basis set, charge, spin state, solvent model, convergence PreprintMolLLMAgent criterion, and computational budget. Materials-agent surveys make a similar argument for atomistic and materials workflows [171]. Molecular dynamics and free-energy workflows. Molecular dy- namics tools move agents from static structures to time-dependent behavior. They support studies of protein flexibility, ligand stabil- ity, solvent effects, conformational transitions, and binding-related dynamics. GROMACS, OpenMM, AmberTools, and LAMMPS pro- vide the computational substrate for many of these workflows [1, 15,24,137]. Free-energy workflows add methods such as FEP, TI, M/PBSA, ABFE, umbrella sampling, and related protocols for ranking candidates. MDCrow [13] automates molecular-dynamics workflows, including setup, execution, and analysis. DynaMate [37] targets autonomous protein and protein-ligand MD workflows, while MDAgent2 [121] studies code generation, execution, eval- uation, and self-correction for MD simulations. ToolMol [179], MADD [125], TRACE [75], modular task-execution agents [102], and MolClaw [174] combine MD or free-energy-like validation with docking, ADMET, and optimization. Evaluation has to con- sider stability, convergence, sampling sufficiency, cost, and whether simulation feedback changes downstream decisions. 2.3.3 Synthesis, characterization, and experimental action tools. Reaction prediction and synthesis feasibility. A molecular can- didate is not actionable until the agent can connect it to feasi- ble synthesis, available starting materials, and executable reac- tion conditions. Reaction-prediction and retrosynthesis tools es- timate whether a proposed molecule can be made, how many steps may be required, which reagents or catalysts are plausible, and whether route constraints match the intended application. AiZynthFinder and ASKCOS are useful reference points for this layer [33,141]. Agentic systems increasingly treat synthesis plan- ning as part of molecular design rather than as post-processing. ChemCrow [91] uses synthesis-oriented tools in chemistry problem solving. Chemist-X [16] focuses on reaction-condition recommen- dation, RETRO-R1 [86] studies agentic retrosynthesis, and Llam- ole [80] integrates inverse molecular design with retrosynthetic planning. ChemActor [177] converts synthesis procedures into structured chemical action sequences. Together, these works move the action space from attractive structures to candidates that can plausibly enter a Design, Make, Test, Analyze cycle. Characterization, spectroscopy, and experimental readout. Closed- loop molecular agents must read experimental evidence, not just propose experiments. Characterization and spectroscopy tools con- vert measurements into structured feedback: NMR assignments, MS or LC-MS peaks, IR or UV spectra, XRD patterns, XANES fea- tures, microscopy images, plots, and assay readouts. The NIST Chemistry WebBook, MassBank, NMRShiftDB, and FDMNES pro- vide representative infrastructure for thermochemical, spectral, NMR, MS, and XANES analysis [36,44,52,59,79,97,129]. This feedback tells the agent whether a reaction succeeded, whether the expected product formed, whether impurities appeared, and whether the next experiment should continue, change, or stop. LLM-RDF [116], ORGANA [22], and robotic ChemAgents [126] demonstrate experimental execution and readout interpretation. Tippy [27,28] specifies corresponding laboratory-facing roles, but its reported evidence does not establish autonomous physical execu- tion. ChemGraph-XANES [35] gives a concrete XANES simulation and analysis workflow for agents, while ChemLabs [156] examines multimodal reasoning in chemistry. The central issue is uncertainty: spectra and curves can support several hypotheses, so the agent should preserve alternatives instead of selecting one explanation too early. Laboratory automation and safety gates. Laboratory automa- tion tools move molecular agents from recommendation to phys- ical action. They include protocol generators, robotic synthesis platforms, liquid-handling systems, cloud laboratories, reaction- execution interfaces, and instrument-control tools. This step cre- ates stronger governance requirements. A failed computational job wastes time; a failed laboratory action can waste material, damage instruments, create unsafe conditions, or violate compli- ance rules. Coscientist [10], LLM-RDF [116], ORGANA [22], and robotic ChemAgents [126] combine execution, characterization, and feedback in physical workflows. AutoLabs [106] instead evalu- ates hardware-ready protocol generation with human verification before execution, while Tippy [27,28] contributes a laboratory- facing architecture. Safety gates can include controlled-chemical checks, protocol validation, resource constraints, and human ap- proval before high-risk actions, as suggested by safety-aware tool use in ChemCrow [91]. 2.3.4 Tool orchestration and evaluation. Workflow orchestration, state passing, and provenance. Molecular agency becomes most visible when tools are linked into workflows. An agent may pass a molecule from SMILES to SDF, convert a struc- ture into docking input, prepare a PDB file for MD, summarize a tra- jectory into stability metrics, match spectra against candidate struc- tures, or translate a reaction plan into a robot-executable protocol. Each transition can lose stereochemistry, protonation state, units, conformers, file-format details, or experimental conditions. Chem- Crow [91] and CACTUS [93] demonstrate early chemistry tool orchestration. ChemHTS [77], ChemHAS [78], CheMatAgent [152], Mozi [14], and MolClaw [174] study more structured forms of tool stacking, tool learning, hierarchical skills, or governed autonomy. Longer-horizon systems such as El Agente [182], ChemGraph [109], ChemGraph-XANES [35], MDCrow [13], DynaMate [37], MDA- gent2 [121], and LLaMP [19] point to the need for state tracking, logging, versioning, intermediate-memory management, and prove- nance records. ToolUniverse points in the same direction for general scientific tooling, where standardized specifications and reusable composition become infrastructure [32]. Failure recovery, cost control, and calibration. Tool use also in- troduces failure modes that are specific to chemistry. A workflow can fail because of invalid SMILES, lost stereochemistry, missing hydrogens, inconsistent protonation states, malformed docking in- puts, DFT non-convergence, unstable MD trajectories, database mismatches, unit errors, unavailable reagents, or unsafe proto- cols. A robust molecular agent should detect these failures, explain them, retry with corrected inputs, switch tools when appropriate, or request human intervention. Tooling-or-Not-Tooling [166] mo- tivates this caution by showing that tools can be harmful when PreprintConferenceâ17, July 2017, Washington, DC, USA used in the wrong setting. ChemHAS [78], TRACE [75], and Mol- Claw [174] suggest that self-correction, action-level experience, and workflow-level skills can improve tool-facing behavior. Computa- tional workflow agents such as El Agente [182], ChemGraph [109], MDCrow [13], DynaMate [37], and MDAgent2 [121] make cost con- trol especially important, since high-cost DFT, MD, or free-energy calculations should not run when a lower-fidelity check is enough. Calibration matters for the same reason. The agent should know when a proxy score is useful for ranking, when higher-fidelity validation is needed, and when the evidence remains inconclusive. Evaluation across the autonomy ladder. Toolbox evaluation should retain a tool-specific view even when the detailed bench- mark taxonomy is consolidated in Section 3. Beyond final-answer accuracy, it should test molecular identity preservation, valid calls, state passing, recovery, provenance, and whether tool evidence improves a later decision. MolViBench [70], MolBench [174], and MatTools [81] expose parts of this process, while L3 evaluation additionally requires protocol executability, hardware compatibility, safety, and faithful interpretation of measurements. 2.4 Learning and Optimization Learning and optimization describe two complementary dimen- sions of improvement in molecular agents. Learning concerns what the agent acquires or updates from experience, includ- ing policies, memories, predictive models, acquisition strategies, tool-use behaviors, and reusable workflows. Optimization con- cerns what component or decision variable is deliberately improved with respect to an objective, ranging from molecular candidates and experimental conditions to action sequences, tool choices, search policies, and the workflow itself. The distinction is therefore not simply between improvement within a task and improvement across trajectories. Optimization may operate within a single trajectory or across many episodes, while learning may occur online during a trajectory or accumulate across repeated tasks and campaigns. The two processes also inter- act: optimization can exploit previously learned knowledge, while learning itself is often driven by an optimization objective. For molecular agents, the important question is not whether a system contains a learning algorithm or repeatedly performs optimization. Rather, the evidence for agency lies in whether ob- servations, evaluations, or accumulated experience change a later consequential scientific decision. The autonomy level then depends on what kind of decision is changed: a fixed computational procedure remains L1; computational evidence that changes a later candidate, plan, tool call, or stopping decision supports L2; physical execution with incorporated experimental feedback supports L3; and accu- mulated evidence that changes objectives, strategy, or scientific agenda across campaigns is required for L4. 2.4.1 Optimization Targets and Search Spaces. Early molecular learning and optimization systems focused primarily on candidate- level decisions, such as generating a molecule or editing a molecular graph. More recent molecular agents broaden the object of opti- mization from the molecular candidate to the discovery process itself. They may optimize which candidate to evaluate, which tool to invoke, which evidence to trust, how computational resources are allocated, when a failed action should be repaired, and whether a multi-step trajectory should continue or stop. Section 2.1 describes molecular representations in detail; here, the relevant issue is the candidate and action spaces they expose to an optimizer. Sequence policies generate or edit strings, whereas graph policies act on atoms, bonds, fragments, or scaffolds. Property- directed SMILES generation in Molecular De Novo Design [103] and ReLeaSE [111] established the proposeâscoreâupdate pattern. REINVENT4 extends this paradigm through transfer learning, re- inforcement learning, curriculum learning, and multi-component scoring [88]. String actions integrate readily with language models but can fail at ring closures, stereochemistry, or syntax. SELFIES, Group SELF- IES, and SAFE reduce some syntactic failures through constraints or chemically meaningful units [18,58,100], although validity does not imply stability or synthesizability. Graph actions instead make molecular topology and local edits explicit. GCPN constructs graphs under validity constraints, whereas MolDQN performs local lead- optimization edits from an existing molecule [165, 181]. Representation determines what can be changed, but it does not determine whether the system is agentic. The same optimization policy may appear inside a fixed human-controlled pipeline or an adaptive agent. What matters is whether feedback from an earlier action alters a later proposal, evaluation strategy, tool choice, or selection decision. 2.4.2Multi-Objective Decision Making. Evaluation provides the ob- jectives and constraints under which optimization proceeds. Early molecular optimization studies often emphasized a single objective, such as QED, LogP, molecular similarity, predicted binding affinity, or docking score. Drug discovery and materials design, however, generally require simultaneous consideration of activity, selectivity, toxicity, ADMET, stability, synthetic accessibility, novelty, diversity, and cost. Because these criteria may conflict, optimization commonly re- lies on composite objectives, constrained optimization, threshold- based filtering, Pareto selection, or combinations of these mech- anisms. DrugEx v2 [87] combines reinforcement learning with Pareto ranking to balance multiple molecular properties rather than maximize a single weighted score. QADD [180] performs it- erative multi-objective reinforcement learning for de novo drug design, while MARS [155] combines Markov chain Monte Carlo with graph editing in a proposeâevaluateâretain loop. Multi-objective decision making is more than assigning a score. It determines which trade-offs are acceptable, which constraints are non-negotiable, and which candidates receive additional computa- tional or experimental resources. Some objectives can be balanced against one another, whereas synthesis feasibility, safety, or exe- cutability may instead be enforced as hard constraints. These mechanisms remain optimization methods rather than evidence of autonomy by themselves. They contribute to an L2 agent only when their evaluations affect a subsequent candidate, edit, tool invocation, resource-allocation decision, or stopping con- dition without an intervening human decision. Likewise, applying an optimizer to an experimental objective does not establish L3 unless physical execution and incorporated experimental feedback are demonstrated. PreprintMolLLMAgent L4 Scientific-agenda discovery loop L3 Feedback-aware physical loop L2 Adaptive computational loop L1 Assistive or fixed workflow accumulated evidence âresearch agenda physical feedback âą no mandatory human gate digital evidence revises the scientific state human or rule closes the loop externally Evidence base domain + cross- campaign memory Formulate candidate questions Prioritize select question âą task âą resources Run campaigns L2 computation + L3 experiment Synthesize compare âą validate âą update Cross-campaignlearning accumulate âą synthesize âą revise agenda new knowledge + questions Experiment goal constraints + safety AGENT / CONTROLLER design or adapt protocol Physical experiment execute Physical evidence execution + measurement Physical result interpret âą validate âą complete Physical-feedback adaptation interpret âą recover âą next action Compute goal constraints AGENT / CONTROLLER candidate âą plan âą tool âą stop Digital environment tools + simulator Digital evidence scores âą predictions âą status Revised state candidate âą hypothesis âą plan Computational adaptation update state + next decision External control human or predefined rule Fixed objective goal + protocol AGENT workflow fixed retrieval âą tools âąreasoning âą answer Answer / result molecular properties / structures External evaluation / update ground truth or predefined rule Figure 4: Hierarchical learning and optimization in molecular LLM agents. Optimization may target candidates, actions, workflows, or experiments, while learning updates policies, memories, models, strategies, and reusable workflows from experience. These mechanisms can operate at multiple autonomy levels: from fixed computational procedures (L1), through feedback-driven computational decisions (L2) and physical experimentation (L3), to cross-campaign learning that revises higher-level scientific objectives or agendas (L4). Most computational optimization relies on surrogate objectives, including property predictors, docking scores, estimated binding affinity, synthetic accessibility, ADMET models, and other low-cost approximations. These signals make repeated evaluation feasible, but they are not equivalent to the scientific outcomes ultimately sought. A sufficiently capable optimizer may therefore exploit weak- nesses in a proxy rather than discover a scientifically useful candi- date. Such failure becomes especially consequential in long-running agentic loops because biased feedback can affect many subsequent decisions. Robust optimization should consequently incorporate va- lidity checks, uncertainty estimates, applicability-domain analysis, diversity constraints, and higher-fidelity confirmation where ap- propriate, while explicitly distinguishing predicted evidence from measured evidence. 2.4.3 Exploration and Acquisition. Optimization also requires de- ciding where additional search or evidence is most valuable. Chem- ical space is extremely large, and useful regions are sparse. An opti- mizer that exploits current rewards too aggressively may converge around a narrow family of high-scoring structures and repeatedly make minor local modifications, reducing scaffold diversity and increasing vulnerability to biased surrogate objectives. Mol-AIR [159] introduces adaptive intrinsic rewards for goal- directed molecular generation, encouraging novelty and broader coverage in addition to target-property performance. Augmented Hill-Climb [136] improves the sample efficiency of REINVENT-like molecular language models, reducing the number of unproductive candidates requiring downstream evaluation. When evaluations are expensive, acquisition itself becomes an optimization problem. Phoenics [40] uses previous observations and predictive uncertainty to recommend candidates or conditions expected to provide either high objective value or useful informa- tion. The constrained latent-space approach of GĂłmez-Bombarelli et al. [34] similarly provides a smoother domain in which to conduct search than direct optimization over discrete strings or graphs. Exploration becomes agentic when accumulated evidence changes where the controller searches, what information it chooses to acquire, how much resource it allocates to an evaluation, or when it decides that further search is no longer worthwhile. This distinction becomes increasingly important as the workflow moves from inexpensive computational scoring toward simulations and physical experiments. 2.4.4 Agent Learning and Workflow Optimization. Molecular agents extend improvement beyond the candidate itself. An agent may learn which tools are reliable, which representations are useful, which failures recur, which search regions are productive, or which workflows succeed under particular objectives. It may then opti- mize later behavior using this acquired information by changing PreprintConferenceâ17, July 2017, Washington, DC, USA its policy, tool sequence, resource allocation, repair strategy, or stopping rule. ReMol [150] combines LLM guidance with reinforcement learn- ing. The language model contributes chemical priors and reasoning signals, while reinforcement learning updates the molecular policy using property feedback. This illustrates how learned knowledge can influence optimization without restricting the LLM to direct molecular-string generation. ChemCRAFT [64] learns tool-use policies from trajectories in chemical sandboxes. MolClaw [174] organizes tool-level, workflow- level, and discipline-level skills for molecular evaluation, screening, and optimization. These systems illustrate learning at the level of agent behavior rather than only at the level of molecular generation. General agent-learning methods provide related mechanisms. Reflexion [123] stores natural-language feedback that can alter behavior in later attempts without parameter updates. Agent Light- ning [90] represents multi-step agent execution as a Markov deci- sion process and applies reinforcement learning to assign credit across trajectories. The existence of memory, reinforcement learning, or trajectory storage nevertheless does not establish scientific agency. Learning is scientifically consequential only when acquired information is reused to change a later decision, and when the effect is demon- strated beyond the episode from which the information was ob- tained. Likewise, workflow optimization requires evidence that feedback changes a consequential candidate, hypothesis, plan, tool choice, resource-allocation decision, or stopping rule. Merely ex- ecuting a predefined sequence of tools, storing a transcript, or correcting tool syntax remains compatible with L1. 2.4.5 Laboratory-in-the-Loop Optimization. Laboratory-in-the- loop optimization replaces or supplements computational proxies with physical observations. The optimization target may be a molecular candidate, reaction condition, formulation, synthesis protocol, or experimental sequence, while learning can update the acquisition policy, predictive model, experimental strategy, or workflow from observed outcomes. Experimental feedback is typically slower, noisier, and more expensive than computational scoring. The agent must therefore balance target performance with information gain, material use, executability, reproducibility, uncertainty, and safety. Bayesian reaction optimization [122] demonstrates how previous observations and uncertainty can guide the selection of subsequent experimental conditions. The method provides an experimental learning and optimization mechanism, although that mechanism alone does not establish an LLM-centered autonomous agent. Self-driving laboratories provide clearer evidence of closed phys- ical loops. AlphaFlow [146] uses reinforcement learning to guide a microfluidic platform in the exploration and optimization of multi- step chemical processes. Autonomous Polymer Synthesis [57] per- forms multi-objective closed-loop optimization for polymer syn- thesis under Pareto trade-offs. In both cases, measured outcomes affect subsequent experimental decisions. LLM-centered systems increasingly connect planning with laboratory execution. Coscientist [10] plans and executes experi- ments through equipment or cloud-laboratory interfaces and uses feedback to interpret measurements or repair execution code. OR- GANA [22] combines agent-generated plans with visual feedback during physical execution. AutoLabs [106] evaluates self-corrected, hardware-ready protocol generation, but a mandatory human verification step before execution prevents the reported workflow from satisfying L3 under our criterion. A stronger form of L3 occurs when a measurement determines another experiment. LLM-RDF [116] uses measured reaction yields to select subsequent reaction conditions. Robotic ChemAgents [126] uses measured catalyst performance to select later compositions for physical validation. The Mobile Robotic Chemist [12] and A- Lab [131] likewise use observed experimental outcomes to choose subsequent physical actions. These cases distinguish broad L3 from iterative L3. Broad L3 requires autonomous physical execution together with incorpo- rated execution or measurement feedback that changes recovery, interpretation, or completion decisions. Iterative L3 requires the stronger condition that a measured result determines a subsequent physical experiment or experimental condition. Human monitoring and emergency stopping do not necessarily reduce the autonomy level, whereas a mandatory approval gate be- fore routine physical execution does. Fixed protocol replay remains L1, and a proposed laboratory architecture without demonstrated physical execution does not establish L3. 2.4.6 Cross-Trajectory and Cross-Campaign Learning. Learning can persist beyond a single trajectory. Across repeated tasks, an agent may accumulate reusable memories, update policies, estimate tool reliability, refine acquisition strategies, or induce workflow-level skills from previous successes and failures. Such cross-trajectory learning can improve later L2 or L3 decisions without necessarily changing the scientific objective itself. Cross-campaign learning is stronger. Here, accumulated evi- dence from completed optimization or experimental campaigns alters higher-level scientific choices, such as which objective to pursue, which hypothesis to investigate, which region of chemical space deserves further study, which experimental strategy should be abandoned, or how future campaigns should be organized. This distinction is important for L4. Reusing a successful work- flow, updating a policy from previous trajectories, or fine-tuning an agent from accumulated experience may constitute learning, but it does not by itself establish L4 autonomy. L4 requires evidence that learning changes a consequential scientific decision at the cam- paign or agenda level rather than merely improving execution of a previously specified objective. 2.4.7What Learning and Optimization Evidence Establishes. Learn- ing and optimization mechanisms should therefore not be mapped directly onto autonomy levels. Reinforcement learning, Bayesian op- timization, memory, multi-objective selection, and workflow adap- tation can all occur at different levels depending on how their outputs affect subsequent decisions. A fixed generator, scorer, optimizer, or workflow remains L1 when iteration is prescribed by a human or fixed procedure. L2 requires computational evidence or learned experience to change a later candidate, hypothesis, plan, tool call, acquisition decision, or stopping condition. L3 additionally requires physical execution together with incorporated execution or measurement feedback; PreprintMolLLMAgent Table 2: Landscape of representative molecular and scientific-agent benchmarks. The table emphasizes evaluated interaction and evidence rather than ranking benchmark-specific scores, which are not directly comparable. BenchmarkScientific scopeAgent interaction or loopPrimary evaluation signal MolViBench [70]358 molecular tasks, 12 workflows, and five difficulty levels Generates executable programs for multi-step molecular workflowsProgram correctness, execution success, and degradation with workflow complexity MolBench [174]Molecular screening, optimization, and end-to-end discovery challenges spanning 8â50+ tool calls Connects filtering, affinity estimation, molecular editing, and workflow execution Subtask quality and end-to-end challenge completion ChemCost [153]1,427 reactions, 2,261 chemicals, and 230,775 price quotes Reasons over reaction components and noisy procurement informationCost-error tolerance, robustness to noise, and component-level attribution MDGym [60]169 molecular simulations and 303 tasks across two MD engines Configures, runs, diagnoses, and repairs simulation workflowsExecutable task success by difficulty, engine, and failure type ChemReason- Bench [172] 7,306 tasks instantiated from 500 organic reactions in six formats Produces or validates ordered, condition-aware, schema-constrained procedure steps Ordering, constraint validation, entity-role grounding, and parseable completion Corral [2]Four chemistry and materials environments: MD, ML, catalysis, and spectroscopy Compares tool-calling and ReAct-style agents in executable expert-designed tasks Task success, tool-use failures, and sensitivity to taskâtool alignment MADE [92]Closed-loop computational materials discovery over chemical systems Proposes and evaluates candidates under a constrained oracle budget, then adapts the search Discovery efficacy, efficiency, and scaling with search-space complexity ScienceAgentBench [17]102 tasks derived from 44 papers across four disciplines Produces executable research code with paper and data contextFull-task and partial success, expert-knowledge dependence, and execution correctness SciAgentBench [120]259 tasks, 1,134 subquestions, and 1,780 scientific tools Selects and composes tools from elementary calls to long workflowsStep and overall success, tool routing, and performance versus interaction horizon SciCode [138]80 scientific coding problems and 338 expert-designed subproblems Implements research algorithms from specifications and intermediate requirements Subproblem and full-problem execution success CORE-Bench [124]270 reproducibility tasks from 90 papers in three disciplines Reproduces published results from code, data, text, and visual artifactsReproduction accuracy across difficulty levels and modalities SciAgentArena [85]Approximately 200 real-world scientific tasks across multiple domains Solves interactive research scenarios with stepwise verification Stepwise task completion and behavior in specified versus open-ended scenarios measurement-selected follow-up experiments provide the stronger iterative form. L4 requires learning across campaigns to influence higher-level objective selection, strategy, hypothesis formation, or scientific agenda. Current evidence therefore spans candidate and workflow opti- mization at L1 and L2, together with bounded physical optimization and learning at L3. Coscientist and ORGANA [10,22] demonstrate feedback-aware physical workflows, while AlphaFlow, autonomous polymer synthesis, LLM-RDF, Robotic ChemAgents, the Mobile Robotic Chemist, and A-Lab [12,57,116,126,131,146] demonstrate measurement-driven experimental iteration. These results show increasingly capable learning and optimization loops, but none by itself establishes the cross-campaign objective selection and scientific-agenda revision required for L4. 3 Evaluation and Benchmarking Evaluation must distinguish component competence from scientific loop closure. Executable-workflow benchmarks can reveal whether an agent constructs valid programs, selects tools, and completes multi-step interactions correctly [70,120]. Closed-loop discovery environments ask a different question: whether observations redi- rect a budgeted search toward better candidates [92]. A chemistry question-answering score therefore cannot establish reliable tool use, while a single end-to-end success rate can hide whether the decisive contribution came from the controller, foundation model, tool, or evaluator. We organize representative benchmarks by the interaction they expose and the scientific state they require the agent to change. Because benchmark-specific scores are not on a common scale, Figure 5 provides one compact, source-faithful result slice for each benchmark in Table 2, rather than a cross-benchmark aggregate. Molecular-dynamics or chemistry subsets are used when a primary source tabulates them. ScienceAgentBench and SciCode include computational chemistry or chemistry problems but do not tabulate model-level chemistry scores, and CORE-Bench has no molecular split; their panels are therefore explicitly marked as overall results. Each panel retains its source metric and must be interpreted locally. Figure 5 supports four observations when the scores are read together with each benchmarkâs design, rather than as a shared leaderboard. Insight 1: benchmark scores answer different scientific questions. ChemReason-Bench and ChemCost evaluate bounded procedural or cost reasoning against specified targets [153,172]. MolViBench, MDGym, Corral, and SciAgentBench instead require executable tool sequences, making routing, intermediate validity, and recovery part of the evaluated object [2,60,70,120]. Sci- enceAgentBench, SciCode, and CORE-Bench extend the horizon to research code or reproducibility artifacts, whereas MADE and the SciAgentArena optimization slice evaluate feedback-guided search under an oracle budget [17,85,92,124,138]. Thus, superfi- cially similar percentages can denote answer accuracy, executable completion, reproduction success, or search efficiency. Insight 2: outcomes are configuration-sensitive, but more tools are not uniformly better. In the reported MolBench slice, MolClaw-C reaches 81.1% accuracy versus 51.4% for vanilla agents and a 45.1% standalone-LLM mean. For matched ChemCost back- bones, ReAct with tools is associated with a CTA@25 change from 2.5â4.1% to 34.6â50.6% [153,174]. Yet SciAgentBenchâs OtherTools ablation scores 21.0% versus 28.6% for the base configuration, and Corral changes the relative ordering of ReAct and tool calling across backbones and single versus chained tasks [2,120]. Framework claims therefore need matched backbone, controller, tool-set, and task conditions. Insight 3: longer or harder execution exposes a reliability gap. SciCode reports 21.2â28.5% subproblem Pass@1 but only 1.5â7.7% on complete main problems [138]. MDGym full success falls from at most 21% on easy tasks to at most 4% on hard tasks, while CORE- Bench drops from 60.0% to 21.5% for GPT-4o and from 44.4% to PreprintConferenceâ17, July 2017, Washington, DC, USA 02040 Pass@1 (%) Claude-T Claude Gemini DeepSeek 39.7 38.6 36.0 34.6 (a) MolViBench - IR 04080 Accuracy (%) MolClaw Vanilla LLM mean 81.1 51.4 45.1 (b) MolBench 02550 CTA@25 (%) Qwen DS-V4 Claude 4.1 4.1 2.5 50.6 49.4 34.6 (c) ChemCost CoTReAct 01020 Full success (%) Claude Codex Qwen GPT-OSS 21 21 2 0 7 4 0 0 4 4 0 0 (d) MDGym EMH 03570 Score GPT-5.2 Gemini Grok-4 DeepSeek KALE-8B 70.3 67.1 66.6 65.2 49.2 (e) ChemReason 036 Acceleration factor MLIP LLM orch. Chem/LLM Chem/div. Random 6.4 5.4 3.9 2.1 1.0 (f) MADE 04080 MD Pass@5 (%) Cl. ReAct Cl. Tool 4o ReAct 4o Tool 57 69 42 19 69 65 72 73 (g) Corral - MD SingleChain 02040 Self-Debug SR* (%) Claude Mistral GPT-4o Llama70B 32.4 23.5 22.6 13.7 34.3 27.5 23.5 16.7 (h) ScienceAgentBench* No KExpert 02040 Success rate (%) Chem Merged Physics Base Tools 35.6 35.2 31.5 28.6 21.0 (i) SciAgentBench - chem 01530 Pass@1* (%) o1-preview Claude GPT-4o DeepSeek 7.7 4.6 1.5 3.1 28.5 26.0 25.0 21.2 (j) SciCode* MainSub 03060 Accuracy* (%) GPT-4o 4o-mini 60.0 44.4 57.8 32.6 21.5 16.3 (k) CORE-Bench* EMH PenLogP QED DRD2 Vals. ScafHop Osim. Celecox. Aripip. C11H24 Median1 task score 0--1 (dark = high) GPT-5.2 Claude-4.6 Gemini-3 Biomni ClaudeCode ToolUniv. CACTUS DELTA ChemTool DrugAgent LIDDiA ChemCrow hatch = N/A (l) SciAgentArena - optimization Figure 5: Compact reported result slices for all 12 benchmarks in Table 2: (a) MolViBench IR [70]; (b) MolBench [174]; (c) ChemCost [153]; (d) MDGym [60]; (e) ChemReason-Bench [172]; (f) MADE [92]; (g) Corral MD [2]; (h) ScienceAgentBench [17]; (i) SciAgentBench chemistry [120]; (j) SciCode [138]; (k) CORE-Bench [124]; and (l) SciAgentArena molecule optimization [85]. Asterisks mark overall results used where no numeric molecular subgroup is published. In (l), darker cells denote higher task scores on the sourceâs 0â1 scale and hatching denotes incompatible agentâtask combinations. Axes retain source metrics and are comparable only within panels. 16.3% for GPT-4o-mini [60,124]. Across these distinct designs, the pattern is consistent with errors accumulating across specification, execution, diagnosis, and repair. Partial credit is informative, but should be paired with end-to-end success and results stratified by horizon or difficulty. Insight 4: the metric must expose efficiency and coverage, not only success. MADE reports acceleration over a random baseline under a fixed oracle budget, so it measures how efficiently feed- back redirects discovery rather than whether a single answer is correct [92]. ScienceAgentBench separately reports partial/full suc- cess and expert-knowledge conditions, revealing that added context does not remove the execution gap in the displayed settings [17]. SciAgentArenaâs heatmap further combines strong, failed, and un- supported agentâtask pairs [85]; a mean over completed cells would therefore hide both specialization and missing coverage. These characteristics also delimit autonomy claims. L1 can use bounded correctness or fixed-workflow completion, whereas L2 needs evidence that an observation changes a candidate, plan, tool choice, or stopping decision, along with loop completion, recovery, cost, and oracle use. L3 additionally requires repeatable physical execution, measurement-conditioned decisions, human interven- tions, and safety violations, consistent with multidimensional self- driving-laboratory evaluation [145]. No benchmark in Table 2 yet establishes a general L3 standard or the cross-campaign agenda revision required for L4. Accordingly, headline scores should be accompanied by task coverage, intermediate validity, end-to-end success, efficiency, reproducibility, operational domain, and manda- tory human gates. 4 Evidence for Scientific Autonomy Levels Molecular LLM agents differ in their scientific capabilities and in the extent to which they independently control a scientific work- flow. Laboratory-autonomy frameworks separate physical process execution from data analysis, interpretation, decision-making, and communication, and they classify systems partly by the decisions that still require a human researcher [47,145]. We follow this em- phasis on decision authority while adapting it to LLM-centered molecular agents. Implementation features such as planning, mem- ory, reinforcement learning, tool use, or multi-agent coordination remain enabling mechanisms: a multi-agent system may require approval at every consequential step, whereas one controller may complete a narrow experimental episode without such a gate. As shown in Figure 2, we define scientific autonomy according to the outermost feedback loop that an agent can reliably close without mandatory human intervention. This is an operational rubric for this survey rather than a claim that the field has converged on one universal scale. The distinction concerns decision responsibility and the source of feedback, not the number of tools, reasoning steps, model components, or repeated trials. PreprintMolLLMAgent Scientific workflows as nested feedback loops. Letídenote one scientific workflow episode executed by agentíŽ. We consider three nested forms of loop closure: í â comp, phys, sci,(1) corresponding respectively to computational, experimental, and scientific-agenda loops. For each loop type í , define í¶ í (íŽ,í)= I Loop í (íŽ,í)â§ í» í (í)= 0 ,(2) whereLoop í (íŽ,í)indicates that the corresponding feedback loop is completed,í» í (í)is the number of mandatory human decision gates inside that loop, and I[·] is the indicator function. A mandatory human gate is an intervention without which the ordinary workflow cannot continue. Human monitoring, retrospec- tive inspection, emergency stopping, and specification of initial constraints do not count as mandatory gates when the agent can otherwise proceed. By contrast, required protocol verification be- fore instrument execution is a gate; AutoLabs explicitly reports such a verification step [106]. The three indicators are defined as follows. âą í¶ comp =1 if the agent uses tool or environment evidence to revise a scientifically consequential state or decision, such as a candidate, hypothesis, plan, tool choice, or stopping rule, without an intermediate human decision. When no physical experiment is controlled, this evidence must come from a digital or simulated environment. Executing a fixed pipeline or repairing only syntax does not satisfy this criterion. âą í¶ phys =1 ifí¶ comp =1 and the agent additionally designs or adapts and executes a physical experiment and incorporates execution or measurement feedback into recovery, state, in- terpretation, or completion without an intermediate human decision. Selecting a subsequent experiment from measure- ments is a stronger iterative form of physical autonomy, but is not required forí¶ phys = 1. âą í¶ sci =1 ifí¶ phys =1 and the agent additionally uses accu- mulated evidence to formulate, prioritize, and pursue new scientific questions, including the autonomous selection of transitions across different scientific task families. These loop types are hierarchical: í¶ sci (íŽ,í) â€ í¶ phys (íŽ,í) â€ í¶ comp (íŽ,í).(3) Scientific-agenda autonomy therefore presupposes the ability to manage the experimental and computational processes required to investigate selected questions. Physical autonomy likewise pre- supposes adaptive planning, analysis, or control. A robot that only replays a fixed protocol does not satisfy eitherí¶ comp orí¶ phys . This criterion separates automation of execution from adaptive decision- making, a distinction also made in laboratory-autonomy frame- works [47]. Autonomy-level definition. Given the nested loop indicators, the episode-level autonomy of agent íŽ is defined as íż(íŽ,í)= 1+í¶ comp (íŽ,í)+í¶ phys (íŽ,í)+í¶ sci (íŽ,í).(4) Because of Eq. (3), íż(íŽ,í) â 1, 2, 3, 4. This produces the following four levels. Table 3: Levels of scientific autonomy for molecular LLM agents. Level Closed loopOperational criterion L1NoneThe agent retrieves information, invokes tools, provides recommendations, or executes a fixed workflow, but evidence does not autonomously revise a consequential scientific decision. L2Adaptive computationalGiven a computational objective, the agent uses digital evidence to revise a candidate, hypothesis, plan, tool choice, or stopping decision within a digital or simulated environment. L3Physical workflowGiven a high-level experimental objective, the agent designs and executes a physical experiment and incorporates execution or measurement feed- back without a mandatory human gate during ordinary operation. L4Scientific agendaThe agent formulates and prioritizes new scientific questions, selects suitable computational and ex- perimental task families, and updates its research agenda from accumulated evidence. The four levels may therefore be summarized as follows: L1 systems remain assistive or fixed; L2 systems adapt a computational scientific state from digital evidence; L3 systems autonomously complete a feedback-aware physical workflow; and L4 systems additionally control the evolution of the scientific agenda. Clarifying the level boundaries. The number of tool calls or agents is not an autonomy criterion. A workflow involving many tools remains at L1 when its sequence is fixed or when outputs do not change a later scientific decision. A relatively simple optimization workflow may qualify as L2 if computational evidence changes the candidate, plan, tool choice, or stopping rule. ReAct-style task completion, format repair, and review-only debate do not qualify by themselves. Physical execution alone is also insufficient for L3. A robotic platform that replays a fixed human-written protocol performs automation. L3 requires the agent to formulate or adapt the ex- perimental procedure and to use physical feedback in recovery, interpretation, or completion. The statement that L3 does not require a second experiment refers only to the number of physical trials, not to the absence of feedback. A single episode can qualify when an observation changes what the agent does or concludes inside that episode. For example, Coscien- tist reads UVâVis spectra to identify an initially unknown physical state and, in a separate integrated experiment, consults hardware documentation and repairs an invalid automation method before successful execution [10]. ORGANA likewise uses visual feedback to guide long-horizon physical plans [22]. Neither example needs a second synthesis to demonstrate feedback-aware physical con- trol. When a measured yield or material property selects another experiment, as in LLM-RDF and robotic ChemAgents [116,126], we label the evidence iterative L3. This distinction is specific to our physical-workflow rubric; SDL taxonomies often reserve their stronger closed-loop categories for systems that also automate ex- periment selection across repeated trials [145]. Generating candidate hypotheses is also insufficient for L4. The agent must evaluate candidate questions, select one to pursue, allo- cate computational or experimental actions, and revise its research PreprintConferenceâ17, July 2017, Washington, DC, USA Table 4: Representative molecular-agent systems classified by the strongest feedback loop demonstrated in their reported evaluation. The examples are illustrative rather than exhaustive; level is assigned from evidence-conditioned behavior, not from the number of tools or agents. SystemScientific settingStrongest demonstrated feedback or actionAssigned level ChemAgent [133]Chemical reasoningUpdates task memory to improve bounded reasoning, without closing an external scientific feedback loopL1: assistive memory CACTUS [93]Chemistry tool useExecutes a predefined tool-mediated problem-solving workflowL1: fixed workflow Chemist-X [16]Reaction conditionsRetrieves evidence and recommends conditions; downstream experimental revision remains externalL1: recommendation ChemCrow [91]General chemistryUses computational chemistry tools and their outputs to revise a tool-mediated solution trajectoryL2: computational loop ChatDrug [84]Molecular editingUses retrieval and domain feedback to revise subsequent molecular editsL2: feedback-driven edit ChemReasoner [128]Catalyst and molecular searchUses quantum-chemical rewards to redirect search over candidate reasoning pathsL2: simulation feedback MDCrow [13]Molecular dynamicsUses execution and analysis feedback to diagnose and repair computational MD workflowsL2: workflow adaptation TRACE [75]Lead optimizationReuses tool trajectories and changes candidates or tool plans from computed evidenceL2: optimization loop Coscientist [10]Chemical experimentation Plans and executes physical experiments and incorporates execution feedback during ordinary operationL3: physical workflow LLM-RDF [116]Robotic chemistryConverts plans into robotic execution and interprets experimental readoutsL3: experiment feedback ORGANA [22]Automated laboratoryGenerates procedures, executes them physically, and uses observed results in completion decisionsL3: experiment feedback AutoLabs [106]Laboratory protocol generationSelf-corrects hardware-ready procedures, followed by required human verification before executionL2: computational workflow No surveyed systemCross-campaign discoveryNo evaluated system demonstrates evidence-conditioned formulation and revision of a scientific agenda across task families L4: evidence gap direction from the evidence. Transitions between task families must be evidence-conditioned rather than fixed entirely in advance. Domain-conditioned autonomy. An autonomy claim is meaning- ful only relative to the domain in which the system operates; exist- ing laboratory frameworks likewise caution that a level does not by itself define the scope of the research being categorized [47]. LetD denote a scientific operational domain specifying supported tasks, tools, environments, instruments, resources, and safety constraints. The system-level autonomy withinD can be defined as íż D (íŽ)= max â : Pr íâŒD [ íż(íŽ,í) â„ â ] â„ í ,(5) where í is a predefined reliability threshold. Safety can be imposed as a separate evaluation constraint: Pr íâŒD [ í(íŽ,í)= 1 ] †íż,(6) whereí(íŽ,í)indicates an unsafe or invalid episode andíżdenotes the maximum acceptable violation rate. Separating Eqs.(5)and(6)is important because autonomy and competence are distinct. A system may be authorized to complete an L3 workflow but do so unreliably, whereas a highly accurate molecular predictor may remain at L1 because it neither determines nor executes subsequent actions. Complementary continuous measures. The discrete level can be accompanied by continuous measures of human involvement and loop reliability. This follows proposals to report autonomy along- side lifetime, throughput, precision, material use, accessible param- eter space, and optimization performance [145]. For example, the mandatory human-intervention rate may be reported as í H = í gate í decision ,(7) whereí gate is the number of mandatory human gates andí decision is the number of consequential workflow decisions. For loop type í , the empirical loop-completion rate is í í = 1 í í âïž í=1 í¶ í (íŽ,í í ).(8) These measures distinguish systems that nominally belong to the same autonomy level but differ substantially in robustness or de- pendence on human intervention. Relationship to enabling technologies. Planning, memory, rein- forcement learning, and multi-agent coordination are orthogonal to the autonomy levels. They are enabling mechanisms whose roles change across levels rather than level-defining properties. At L1, memory may support information retrieval and personal- ized assistance. At L2, it may preserve state across long computa- tional workflows and support recovery from failed tool calls. At L3, it may maintain experimental context across physical iterations. At L4, it may organize evidence accumulated over multiple projects and support long-term research-agenda formation. The same principle applies to reinforcement learning and multi- agent coordination. Their contribution to autonomy should be as- sessed by whether they enable the system to close a broader feed- back loop, rather than by their mere presence in the architecture. 5 Safety and Challenges The architectural components in Section 2 expose distinct but cou- pled failure modes. Perception can corrupt molecular identity, the agent core can select an unsupported action, a tool can return unre- liable evidence, and optimization can amplify weaknesses in any of the preceding components. In this section, we provide an overview of the safety concerns and then organize the open challenges along the same four-part architecture which can be summarized as Fig- ure 6. 5.1 Safety and Governance Boundaries The risk posed by a molecular agent is determined by its operational authority, the reversibility of its actions, the scope of substances and equipment it can affect, and the consequences of an erroneous decision. The self-driving-laboratory safety literature accordingly treats safety as a property of interactions among software, hard- ware, materials, people, and operating procedures rather than of the language model alone [63]. For instance, read-only retrieval and identity checks can be assigned to a control surface narrower PreprintMolLLMAgent Input Action Feedback LoopReasoning Data & Representation Issues (Input Layer) Cognitive & Framework Deficiencies (Brain Layer) Unreliable Internal Logic & Fragmented Memory Weak Orchestration & Long-Term Safety Awareness Scientific Alignment & Optimization Bottlenecks (Goal Layer) Misaligned & Conflicting Metrics Workflow & Action Space Imbalances Tool Execution & Environment Friction (Action Layer) Inadequate Oversight in Tool Usage Lack of Environment- Grounded Benchmarks Information Loss across Conversions Lack of Standardization in Data Passing C1=C=C2=C1C(C)OC2 O Stereochemistry Altered C1=C=C2=C1C(C)OC2 O Graph Database Query Time Immediate Gain Long-Term Catastrophic Risk Data is Not Shared Local Memory for Task 1 Local Memory for Task 2 Potency Toxicity Novelty synthesizability Wasted MaterialsDamaged Instruments Unsafe ConditionsViolated Compliance C1C1 C1CCNCC1 Unexpected Change C1C2 Invalid SMILE-Based Optimization Small action space (sequence length), but high brittleness Task-Level Evaluation MoleculeNet [134] GuacaMol [10] MOSES [96] TDC [43] Process Evaluation CheMatAgent [133] MolViBench [61] MolBench [153] MatTools [69] Environment- Grounded Benchmarks Tool Trajectories Evidence Grounding Decision Value Auditability Figure 6: Challenges and bottlenecks in the lifecycle of a molecular agent. The diagram maps vulnerabilities across perception, controller design, tool and environment inter- action, and optimization; arrows indicate how an upstream error can propagate into more consequential system-level behavior as action authority increases. than purchasing, synthesis planning, or instrument control. Fur- thermore, the same tool may also require different permissions across operational domains. The safety concerns raised by molecular discovery agents can be classified into three broad categories: intrinsic system limitations, extrinsic human-induced threats, and physical-world operational risks. At the intrinsic level, these agents rely on reasoning, planning, and memory modules that are prone to instability or insufficiency, which can lead to hallucinations or erroneous predictions of molec- ular properties. Such inaccuracies introduced during early-stage information retrieval can cascade into catastrophic failures through synthesis planning and robotic execution [14,41]. When such erro- neous information guides experimental workflows, it can resulting in resource waste, generation of hazardous byproducts, and po- tential laboratory incidents. Moreover, large language models are known to struggle with long-horizon planning and complex reason- ing tasks, impairing the agentâs ability to anticipate downstream consequences or recognize critical safety checkpoints in multi-step synthetic routes [134]. On the humanâagent interface, extrinsic safety is predominantly challenged by the diversity of user intent. Instructions may range from benign requests to deliberately malicious prompts. Adversarial inputs and jailbreak techniques can subvert the agentâs safety align- ment, compelling it to propose synthetic pathways for hazardous compounds. Dual use is a concrete molecular design concern, the same capabilities that enable beneficial drug discovery can be re- purposed to design highly toxic compounds, controlled substances, or chemical weapons [143]. However, the most consequential safety challenge emerges when molecular discovery agents are integrated with automated labora- tory platforms. In such scenario, digital decision errors translate directly into physical consequences. A single erroneous command affecting reagent sequencing, temperature control, or hazardous ma- terial handling can trigger chemical spills, fires, explosions, or per- sonal injury [14]. This autonomy risk fundamentally distinguishes molecular discovery agents from general-purpose large language models. Their capacity for direct material manipulation demands rigorous risk control mechanisms, sustained human oversight, and a safety evaluation that is specifically tailored to autonomous mol- ecule discovery. For the systems surveyed here, we recommend a governance contract that records the verified molecular identity, allowed tools and resources, approval gates, stopping and abort conditions, and an auditable trace of inputs, intermediate states, tool versions, out- puts, and overrides. This follows the lifecycle view of the NIST AI Risk Management Framework, whose core functions are to govern, map, measure, and manage risk [132]. Chemistry-specific imple- mentations include safety tools in ChemCrow [91], governed skills in Mozi [14], and bounded laboratory interfaces in Coscientist [10]. Evaluation metrics should then capture not only task success but also invalid or unsafe episodes, blocked or escalated actions, budget violations, recovery events, and human interventions. These are proposed reporting requirements for this survey, not claims that any single control stack is sufficient for every laboratory. 5.2 Challenges in Molecular Perception Molecular perception must preserve chemical identity while trans- lating among representations selected for different tasks. This re- quirement becomes harder as agents move from human-checked L1 interactions to long, autonomous, and multimodal trajectories at L2 and above. Representation fidelity across conversions. No single substrate exposes every relevant molecular property. SMILES and SELFIES are convenient for generation, graphs expose topology, 3D struc- tures support physical reasoning, and images or spectra connect the agent to literature and experiments. Routing among these sub- strates can silently alter stereochemistry, protonation, tautomeric state, atom mapping, conformers, or measurement context. Struc- tured formats such as MolJSON and MoleCode make connectivity more explicit [117,157], but future agents still need round-trip val- idation, invariant checks, and persistent identity tracking at every representation boundary. Adaptive representation and token granularity. Representation choice should depend on the next scientific action rather than on a fixed input format. A compact string may be sufficient for data- base lookup, whereas exact editing, docking, or spectroscopy may require graphs, coordinates, or multimodal state. Tokenization fur- ther determines whether edits correspond to characters, atoms, fragments, or structure-aware units. The open problem is to learn when the current representation is inadequate and to switch sub- strates without losing information, exceeding the context budget, or obscuring the action from human review. Uncertainty-aware feedback perception. Tool outputs and experi- mental observations are not self-interpreting facts. Docking scores, property predictions, spectra, and assay measurements depend on model assumptions, units, conditions, and domains of validity. An PreprintConferenceâ17, July 2017, Washington, DC, USA agent must retain these qualifications when converting an obser- vation into its next state. Otherwise, an ambiguous experimental signal or an out-of-domain prediction can be collapsed into false certainty and propagated through later decisions. Perception bench- marks should therefore test identity preservation, unit and condi- tion tracking, uncertainty retention, and recovery from conflicting multimodal evidence together with structure parsing accuracy. 5.3 Challenges in the Agent Framework Although molecular agent design has progressed from tool-using dialogue to workflow control, reliable autonomous discovery re- mains unresolved. Open problems concern verifiability, memory, orchestration, evaluation, and safety. Verifiable and molecule-native reasoning. Current agents can gen- erate structured tool calls, executable code, protocols, and action traces, but they do not yet provide machine-checked chemical proofs or formal constraint certificates [69]. ChemActor represents synthesis procedures as structured chemical actions [177], but for- mal verification of molecule edits, synthesis constraints, safety rules, and protocol preconditions remains open. The reasoning substrate is also unresolved. Text and SMILES are interpretable but lossy, while graph, 3D, and latent states better preserve molecular struc- ture but are harder to audit. MolLingo [98] and LatentChem [164] illustrate this tension between molecule-native expressiveness and interpretability. Persistent memory and reliable reflection. Most memory remains local to one task, session, or campaign. ChemAgent [133] uses self- updating memories, TRACE [75] stores tool trajectories, and Aug- mented Memory [38] uses experience replay. L4 discovery would require cross-task scientific memory for failures, provenance, un- certainty, design rules, and reusable skills. Reflection faces a similar limitation. Systems such as ChatDrug [84], AgentDrug [61], Probe- Before-You-Edit [160], and MT-Mol [54] rely on external feedback, but high-autonomy agents often face delayed, noisy, or missing ground truth. Future evaluation of reflection must therefore con- sider calibration, stopping rules, and when the agent should defer to simulation, experiment, or human review. Adaptive orchestration and evaluation. More tools, memory, and agents are not always better. Tooling-or-Not-Tooling shows that tool use can help some chemistry tasks while hurting others [166]. Orchestration should be adaptive: the controller must decide when to answer directly, retrieve evidence, call tools, coordinate special- ists, or request human approval. Evaluation is also underdeveloped. Most benchmarks measure final task scores rather than controller behavior, such as reasoning traces, tool routing, planning under budget, memory reuse, reflection after failure, and multi-agent co- ordination. ChemLabs [156] and Tooling-or-Not-Tooling [166] are useful steps, but environment-grounded molecular-agent bench- marks remain limited. Safety and governance. The controller-level problem is to enforce the boundaries defined above at the point of decision making. A final text filter cannot validate molecular identity, instrument state, or whether a requested action lies inside the approved domain. Dual-use molecular generation [143] and the expanding scope of self-driving laboratories [63] motivate structure-aware screening, permissioned execution, traceable decisions, and escalation when evidence or authority is insufficient. The open research question is how to assess these controls without relying solely on refusal rate as the safety metric. 5.4 Challenges in Molecular Toolboxes Tools ground molecular agents in external evidence and executable actions, but they also define the operational boundary within which an agent can fail. A reliable toolbox must make capabilities, assump- tions, costs, and hazards visible to the controller. Semantic interoperability and state passing. Chemistry tools use heterogeneous identifiers, file formats, units, parameter conven- tions, and software environments. Ad hoc wrappers hide these differences but rarely guarantee that a state produced by one tool is valid input to the next. Shared, typed interfaces are needed for molecular identity, experimental conditions, uncertainty, prove- nance, and pre- and postconditions. General scientific tool infras- tructures such as ToolUniverse [32] provide a useful foundation, but molecular workflows additionally require chemistry-aware schemas and validation across databases, simulators, synthesis plan- ners, instruments, and robots. Reliability, calibration, and failure recovery. Tool access does not guarantee trustworthy grounding. Databases can disagree, predic- tors can be out of domain, simulations can fail to converge, and laboratory interfaces can return incomplete observations. Agents must detect invalid inputs, timeouts, numerical failures, inconsis- tent outputs, and low confidence before using a result to revise the plan. Because tool augmentation can either help or hurt depend- ing on the task [166], evaluation should measure tool selection, input construction, output interpretation, fallback behavior, and calibration in addition to final-answer accuracy. Reproducibility, cost, and governed execution. Long molecular workflows require versioned tools, recorded parameters, environ- ment metadata, and traceable state transitions so that a result can be reproduced rather than merely narrated. The controller must also trade off information value against latency, compute, assay cost, and instrument access. For synthesis or laboratory action, these resource policies must connect to the run-level permissions, abort conditions, and audit record defined in the safety governance subsection. The toolbox-specific challenge is implementing those controls across heterogeneous software and hardware interfaces while keeping execution interruptible and confined to the approved operational domain. 5.5 Challenges in Molecular Optimization Molecular optimization tests whether an LLM-based molecular agent can improve candidates under scientific constraints, beyond generating valid structures. Agents must search a vast chemical space under noisy feedback, incomplete constraints, and conflicting objectives. Open problems lie in reward design, multi-objective search, action control, and workflow-level feedback. Proxy objectives and reward hacking. Most optimization systems rely on surrogate objectives, such as QED, penalized LogP, docking PreprintMolLLMAgent scores, predicted binding affinity, synthetic accessibility, and AD- MET predictors. Early reinforcement-learning methods showed that molecular generators can be guided by property rewards [103,111], while REINVENT4 supports richer multi-component scoring func- tions [88]. These scores provide fast feedback, but they remain proxies for the real scientific goal. An agent may exploit the evalu- ator and produce molecules that score well but are unstable, toxic, difficult to synthesize, or outside the predictorâs reliable domain. The challenge is to design rewards that consider validity, synthesiz- ability, uncertainty, diversity, and experimental plausibility rather than optimize a single scalar score. Multi-objective trade-offs and local optima. Real molecular de- sign usually involves conflicting objectives, where improving one property may weaken another. For example, higher potency may increase toxicity, while higher novelty may reduce synthesizability. DrugEx v2 [87], QADD [180], and MARS [155] reflect the move from single-objective optimization toward Pareto ranking, itera- tive multi-objective search, and graph-based editing. For agent- based optimization, premature exploitation is another concern. Once an agent finds a high-scoring scaffold, it may keep making small local edits around it rather than explore new chemical re- gions. Exploration-oriented methods such as Mol-AIR [159] and Augmented Hill-Climb [136] address part of this problem through novelty, coverage, or more efficient search. Future agents need better mechanisms for Pareto-aware optimization, diversity preser- vation, and control of exploration and exploitation. Search efficiency and action-space control. Chemical space is too large for blind trial and error, especially when evaluation requires docking, molecular dynamics, DFT, retrosynthesis analysis, or lab- oratory validation. SMILES-based optimization is easy to combine with sequence models [88,103,111], but small token-level edits can invalidate a molecule or change it unexpectedly. Graph-based meth- ods such as GCPN [165] and MolDQN [181] make actions more chemically explicit by operating on atoms, bonds, and graph edits. However, they also introduce larger and more constrained action spaces. The challenge is to define actions that are expressive enough for discovery, constrained enough to avoid invalid chemistry, and efficient enough for multi-round optimization. Workflow-level feedback and credit assignment. Molecular opti- mization increasingly operates as a multi-step workflow rather than a single generation step. LLM-guided and tool-using sys- tems such as ReMol [150], ChemCrow [91], ChemCRAFT [64], and MolClaw [174] show this shift from molecular generators toward broader agentic pipelines. When optimization fails, the cause may lie in generation, tool selection, reward modeling, planning, or molecular format conversion. This creates a workflow-level credit assignment problem. Evaluation should therefore measure final molecular scores together with multi-round improvement, tool-use efficiency, failure recovery, provenance, and robustness to unreli- able feedback. 6 Future Directions The open problems above become more useful when expressed as testable milestones rather than as another list of missing capabilities. We propose a progression in which each stage produces artifacts and measurements needed to justify the next expansion of authority. Milestone 1: a verifiable molecular-state contract. Near-term sys- tems should publish typed schemas for molecular identity, struc- ture, conditions, units, uncertainty, provenance, and permitted edits, together with round-trip conversion tests across strings, graphs, coordinates, images, spectra, and experimental records. MolJSON, MoleCode, and MolLingo [98,117,157] illustrate current attempts to make molecular structure more explicit or molecule-native for LLM reasoning. The milestone is reached when declared conver- sions preserve identity and required metadata on a public test suite, invalid states are rejected before tool execution, and every down- stream result can be traced to its input representation. This is a concrete prerequisite for attributing a later failure to the controller rather than to silent state corruption. Milestone 2: reproducible and governed L2 loops. The next tar- get is not a larger tool catalogue but a computational controller whose trajectories can be replayed and compared under the same evidence and resource budget. It should record tool versions, in- termediate states, uncertainty, stopping decisions, and the reason for each escalation. Persistent memory can then be evaluated by withholding prior trajectories and measuring whether validated experience improves a new run without propagating stale or incor- rect evidence. MolViBench and SciAgentBench expose multi-step tool execution [70,120], while MADE supplies a budgeted closed- loop discovery setting [92]. A credible L2 milestone should report loop completion, improvement over rounds, recovery, cost, repro- ducibility, and mandatory human gates against fixed-workflow and no-memory baselines. Milestone 3: bounded, safety-evaluated physical episodes. An L2 controller should progress to L3 only on a declared experimental domain with validated protocols, instrument limits, permissions, abort conditions, and measurement-quality checks. Coscientist and ORGANA provide primary examples of LLM-centered planning con- nected to physical execution and feedback [10,22]; safety work on self-driving laboratories motivates evaluating the whole softwareâ hardwareâmaterial system [63]. The milestone is repeatable com- pletion of the same bounded episode across multiple trials while reporting execution success, measurement validity, recovery, hu- man intervention, unsafe or invalid episodes, and resource use. A successful demonstration should also state what remains outside the operational domain. Milestone 4: reliable iterative L3 campaigns. The stronger experi- mental target is a loop in which a measurement selects or changes a subsequent physical experiment. LLM-RDF and robotic ChemA- gents demonstrate this pattern using measured yield or catalyst performance [116,126]. Future work should compare the agent with fixed designs and established acquisition policies under matched budgets, replicate selected measurements, and report calibration, sample efficiency, stopping behavior, and recovery from failed or ambiguous readouts. This milestone separates an isolated automa- tion success from an experimental policy that learns reliably during a campaign. PreprintConferenceâ17, July 2017, Washington, DC, USA Milestone 5: evidence for cross-campaign scientific autonomy. L4 should remain an evidence standard rather than an aspirational label. A candidate system would need to preserve evidence across completed campaigns, formulate and prioritize a new question, choose among different computational and experimental task fami- lies, allocate resources, and revise the research agenda when results contradict its assumptions. Evaluation would require multiple task families, prospective rather than retrospective trials, independent scientific review, and evidence that the selected question and result are not recoverable from a fixed predefined workflow. Until such studies exist, progress is better described through the measurable L2 and L3 milestones above than through broad claims of autonomous discovery. 7 Conclusion This survey developed a conceptual framework for molecular LLM agents from two complementary perspectives. The architectural view connects molecular representation and perception, an LLM- centered agent framework, domain-specific toolboxes, and learning and optimization. The scientific-autonomy view classifies agents by the outermost feedback loop they can reliably close without manda- tory human intervention, from L1 assistive or fixed workflows to L4 scientific-agenda agents. Together, these views connect system design with demonstrated decision authority across computational and physical molecular workflows. Our analysis indicates that planning, tool use, stronger language models, or multi-agent coordination alone do not establish scientific autonomy. Progress also requires chemically faithful perception, grounded and verifiable tool use, feedback-aware decision making, persistent provenance, realistic optimization, and safety within an explicit operational domain. Future evaluation should therefore con- sider task success together with loop completion, uncertainty, cost, reproducibility, human involvement, and safety, helping molecular agents advance from useful assistants toward trustworthy partners in molecular discovery. References [1]Mark James Abraham, Teemu Murtola, Roland Schulz, SzilĂĄrd PĂĄl, Jeremy C Smith, Berk Hess, and Erik Lindahl. 2015. GROMACS: High performance molec- ular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 1 (2015), 19â25. [2] Nawaf Alampara, Martiño RĂos-GarcĂa, Chandan Gupta, Sajid Mannan, Santiago Miret, N. M. Anoop Krishnan, and Kevin Maik Jablonka. 2025. Task Alignment Outweighs Framework Choice in Scientific LLM Agents. In NeurIPS 2025 Work- shop on AI for Accelerated Materials Design. https://openreview.net/forum?id= 7cbwuA5k0T [3]Mehrad Ansari and Seyed Mohamad Moosavi. 2024. Agent-based learning of materials datasets from the scientific literature. Digital Discovery 3, 12 (2024), 2607â2617. [4]Mehrad Ansari, Jeffrey Watchorn, Carla E. Brown, and Joseph S. Brown. 2024.dZiner: Rational Inverse Design of Materials with AI Agents. arXiv:2410.03963 [physics.chem-ph] https://arxiv.org/abs/2410.03963 [5]Christoph Bannwarth, Eike Caldeweyher, Sebastian Ehlert, Andreas Hansen, Philipp Pracht, Jakob Seibert, Sebastian Spicher, and Stefan Grimme. 2021. Ex- tended tight-binding quantum chemistry methods. Wiley Interdisciplinary Reviews: Computational Molecular Science 11, 2 (2021), e1493. [6]Christoph Bannwarth, Sebastian Ehlert, and Stefan Grimme. 2019. GFN2- xTBâAn accurate and broadly parametrized self-consistent tight-binding quan- tum chemical method with multipole electrostatics and density-dependent dispersion contributions. Journal of chemical theory and computation 15, 3 (2019), 1652â1671. [7]Aaditya Baranwal, Akshaj Gupta, Yogesh S. Rawat, and Shruti Vyas. 2026. MolSight: Molecular property prediction with images. arXiv preprint arXiv:2605.10157 (2026). [8]A PatrĂcia Bento, Anne Hersey, Eloy FĂ©lix, Greg Landrum, Anna Gaulton, Francis Atkinson, Louisa J Bellis, Marleen De Veij, and Andrew R Leach. 2020. An open source chemical structure curation pipeline using RDKit. Journal of cheminformatics 12, 1 (2020), 51. [9]Helen M Berman, John Westbrook, Zukang Feng, Gary Gilliland, Talapady N Bhat, Helge Weissig, Ilya N Shindyalov, and Philip E Bourne. 2000. The protein data bank. Nucleic acids research 28, 1 (2000), 235â242. [10]Daniil A Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. 2023. Au- tonomous chemical research with large language models. Nature 624, 7992 (2023), 570â578. doi:10.1038/s41586-023-06792-0 [11]Nathan Brown, Marwin Fiscato, Marwin H. S. Segler, and Alain C. Vaucher. 2019. GuacaMol: benchmarking models for de novo molecular design. Journal of Chemical Information and Modeling 59, 3 (2019), 1096â1108. [12]Benjamin Burger, Phillip M. Maffettone, Vladimir V. Gusev, Catherine M. Aitchi- son, Yang Bai, Xiaoyan Wang, Xiaobo Li, Ben M. Alston, Buyi Li, Rob Clowes, Nicola Rankin, Brandon Harris, Reiner S. Sprick, and Andrew I. Cooper. 2020. A mobile robotic chemist. Nature 583 (2020), 237â241. doi:10.1038/s41586-020- 2442-2 [13]Quintina Campbell, Sam Cox, Jorge Medina, Brittany Watterson, and Andrew D White. 2026. Mdcrow: Automating molecular dynamics workflows with large language models. Machine Learning: Science and Technology 7, 2 (2026), 025037. doi:10.1088/2632-2153/ae4b07 [14]He Cao, Siyu Liu, Fan Zhang, Zijing Liu, Hao Li, Bin Feng, Shengyuan Bai, Leqing Chen, Kai Xie, and Yu Li. 2026. Mozi: Governed autonomy for drug discovery llm agents. arXiv preprint arXiv:2603.03655 (2026). https://arxiv.org/abs/2603.03655 [15]David A Case, Hasan Metin Aktulga, Kellon Belfon, David S Cerutti, G AndrĂ©s Cisneros, VinĂcius Wilian D Cruzeiro, Negin Forouzesh, Timothy J Giese, An- dreas W Gotz, Holger Gohlke, et al.2023. AmberTools. Journal of chemical information and modeling 63, 20 (2023), 6183â6191. [16] Kexin Chen, Jiamin Lu, Junyou Li, Xiaoran Yang, Yuyang Du, Kunyi Wang, Qiannuan Shi, Jiahui Yu, Lanqing Li, Jiezhong Qiu, et al.2023. Chemist-X: Large language model-empowered agent for reaction condition recommendation in chemical synthesis. arXiv preprint arXiv:2311.10776 (2023). doi:10.48550/arXiv. 2311.10776 [17] Ziru Chen, Shijie Chen, Yuting Ning, Qianheng Zhang, Boshi Wang, Botao Yu, Yifei Li, Zeyi Liao, Chen Wei, Zitong Lu, Vishal Dey, Mingyi Xue, Frazier N. Baker, Benjamin Burns, Daniel Adu-Ampratwum, Xuhui Huang, Xia Ning, Song Gao, Yu Su, and Huan Sun. 2025. ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery. In International Confer- ence on Learning Representations. https://proceedings.iclr.c/paper_files/paper/ 2025/hash/f12b4df26344f3be803c06b555252efe-Abstract-Conference.html [18]Austin H. Cheng, Andy Cai, Santiago Miret, Gustavo Malkomes, Mariano Phielipp, and AlĂĄn Aspuru-Guzik. 2023. Group SELFIES: a robust fragment- based molecular string representation. Digital Discovery 2, 3 (2023), 748â758. [19]Yuan Chiang, Elvis Hsieh, Chia-Hong Chou, and Janosh Riebesell. 2025. LLaMP: Large Language Model Made Powerful for High-fidelity Materials Knowledge Retrieval. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 25189â25221. doi:10.18653/v1/2025.emnlp-main.1280 [20]Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. 2020. Chem- BERTa: Large-scale self-supervised pretraining for molecular property predic- tion. arXiv preprint arXiv:2010.09885 (2020). [21]Gabriele Corso, Hannes StĂ€rk, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. 2022. Diffdock: Diffusion steps, twists, and turns for molecular docking. arXiv preprint arXiv:2210.01776 (2022). [22] Kourosh Darvish, Marta Skreta, Yuchi Zhao, Naruki Yoshikawa, Sagnik Som, Miroslav Bogdanovic, Yang Cao, Han Hao, Haoping Xu, AlĂĄn Aspuru-Guzik, et al. 2025. ORGANA: A robotic assistant for automated chemistry experimentation and characterization. Matter 8, 2 (2025), 101897. doi:10.1016/j.matt.2024.10.015 [23]Vishal Dey, Xiao Hu, and Xia Ning. 2025. GeLLM3O: Generalizing Large Lan- guage Models for Multi-property Molecule Optimization. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Moham- mad Taher Pilehvar (Eds.). Association for Computational Linguistics, Vienna, Austria, 25192â25221. doi:10.18653/v1/2025.acl-long.1225 [24]Peter Eastman, Jason Swails, John D Chodera, Robert T McGibbon, Yutong Zhao, Kyle A Beauchamp, Lee-Ping Wang, Andrew C Simmonett, Matthew P Harrigan, Chaya D Stern, et al.2017. OpenMM 7: Rapid development of high performance algorithms for molecular dynamics. PLoS computational biology 13, 7 (2017), e1005659. [25]Carl Edwards, Chi Han, Gawon Lee, Thao Nguyen, Bowen Jin, Chetan Kumar Prasad, Sara SzymkuÄ, Bartosz A Grzybowski, Ying Diao, Jiawei Han, et al.2025. mclm: A function-infused and synthesis-friendly modular chemical language model. arXiv e-prints (2025), arXivâ2505. [26]Carl Edwards, Tuan Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji. 2022. Translation between molecules and natural language. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 375â413. PreprintMolLLMAgent [27]Yao Fehlis, Charles Crain, Aidan Jensen, Michael Watson, James Juhasz, Paul Mandel, Betty Liu, Shawn Mahon, Daren Wilson, Nick Lynch-Jonely, et al.2025. Accelerating drug discovery through agentic ai: A multi-agent approach to laboratory automation in the dmta cycle. arXiv preprint arXiv:2507.09023 (2025). [28]Yao Fehlis, Charles Crain, Aidan Jensen, Michael Watson, James Juhasz, Paul Mandel, Betty Liu, Shawn Mahon, Daren Wilson, Nick Lynch-Jonely, et al.2025. Technical Implementation of Tippy: Multi-Agent Architecture and System De- sign for Drug Discovery Laboratory Automation. arXiv preprint arXiv:2507.17852 (2025). [29]M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E. Scuseria, M. A. Robb, J. R. Cheeseman, G. Scalmani, V. Barone, G. A. Petersson, H. Nakatsuji, X. Li, M. Caricato, A. V. Marenich, J. Bloino, B. G. Janesko, R. Gomperts, B. Mennucci, H. P. Hratchian, J. V. Ortiz, A. F. Izmaylov, J. L. Sonnenberg, D. Williams-Young, F. Ding, F. Lipparini, F. Egidi, J. Goings, B. Peng, A. Petrone, T. Henderson, D. Ranasinghe, V. G. Zakrzewski, J. Gao, N. Rega, G. Zheng, W. Liang, M. Hada, M. Ehara, K. Toyota, R. Fukuda, J. Hasegawa, M. Ishida, T. Nakajima, Y. Honda, O. Kitao, H. Nakai, T. Vreven, K. Throssell, J. A. Montgomery, Jr., J. E. Peralta, F. Ogliaro, M. J. Bearpark, J. J. Heyd, E. N. Brothers, K. N. Kudin, V. N. Staroverov, T. A. Keith, R. Kobayashi, J. Normand, K. Raghavachari, A. P. Rendell, J. C. Burant, S. S. Iyengar, J. Tomasi, M. Cossi, J. M. Millam, M. Klene, C. Adamo, R. Cammi, J. W. Ochterski, R. L. Martin, K. Morokuma, O. Farkas, J. B. Foresman, and D. J. Fox. 2016. Gaussian~16 Revision C.01. Gaussian Inc. Wallingford CT. [30]Bowen Gao, Yanwen Huang, Yiqiao Liu, Wenxuan Xie, Wei-Ying Ma, Ya-Qin Zhang, and Yanyan Lan. 2025. Pharmagents: Building a virtual pharma with large language model agents. arXiv preprint arXiv:2503.22164 (2025). [31] Bowen Gao, Yanwen Huang, Yiqiao Liu, Wenxuan Xie, Wei-Ying Ma, Ya-Qin Zhang, and Yanyan Lan. 2025. Pushing the boundaries of structure-based drug design through collaboration with large language models. arXiv preprint arXiv:2503.01376 (2025). [32] Shanghua Gao, Richard Zhu, Pengwei Sui, Zhenglun Kong, Sufian Aldogom, Yepeng Huang, Ayush Noori, Reza Shamji, Krishna Parvataneni, Theodoros Tsiligkaridis, et al.2025. Democratizing AI scientists using ToolUniverse. arXiv preprint arXiv:2509.23426 (2025). [33]Samuel Genheden, Amol Thakkar, Veronika ChadimovĂĄ, Jean-Louis Reymond, Ola Engkvist, and Esben Bjerrum. 2020. AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. Journal of cheminformatics 12, 1 (2020), 70. [34]Rafael GĂłmez-Bombarelli, Jennifer N. Wei, David Duvenaud, JosĂ© Miguel HernĂĄndez-Lobato, BenjamĂn SĂĄnchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and AlĂĄn Aspuru- Guzik. 2018. Automatic chemical design using a data-driven continuous repre- sentation of molecules. ACS Central Science 4, 2 (2018), 268â276. doi:10.1021/ acscentsci.7b00572 [35] Vitor F Grizzi, Thang Duc Pham, Luke N Pretzie, Jiayi Xu, Murat Keceli, and Cong Liu. 2026. ChemGraph-XANES: An Agentic Framework for XANES Simulation and Analysis. arXiv preprint arXiv:2604.16205 (2026). [36]Sergey A Guda, Alexander A Guda, Mikhail A Soldatov, Kirill A Lomachenko, Aram L Bugaev, Carlo Lamberti, Wojciech Gawelda, Christian Bressler, Grig- ory Smolentsev, Alexander V Soldatov, et al.2015. Optimized finite difference method for the full-potential XANES simulations: Application to molecular ad- sorption geometries in MOFs and metalâligand intersystem crossing transients. Journal of chemical theory and computation 11, 9 (2015), 4512â4521. [37]SalomĂ© Guilbert, Cassandra Masschelein, Jeremy Goumaz, Bohdan Naida, and Philippe Schwaller. 2025. DynaMate: An Autonomous Agent for Protein-Ligand Molecular Dynamics Simulations. arXiv preprint arXiv:2512.10034 (2025). [38] Jeff Guo and Philippe Schwaller. 2024. Augmented memory: sample-efficient generative molecular design with reinforcement learning. Jacs Au 4, 6 (2024), 2160â2172. [39]Shuhan Guo, Yatao Bian, Ruibing Wang, Nan Yin, Zhen Wang, and Quanming Yao. 2024. UniMoT: Unified molecule-text language model with discrete token representation. arXiv preprint arXiv:2408.00863 (2024). [40]Florian HĂ€se, LoĂŻc M. Roch, and AlĂĄn Aspuru-Guzik. 2018. Phoenics: A Bayesian optimizer for chemistry. ACS Central Science 4, 9 (2018), 1134â1145. doi:10.1021/ acscentsci.8b00307 [41]Jiyan He, Weitao Feng, Yaosen Min, Jingwei Yi, Kunsheng Tang, Shuai Li, Jie Zhang, Kejiang Chen, Wenbo Zhou, Xing Xie, Weiming Zhang, Nenghai Yu, and Shuxin Zheng. 2023. Control Risk for Potential Misuse of Artificial Intelligence in Science. arXiv:2312.06632 [cs.AI] https://arxiv.org/abs/2312.06632 [42] Stephen R. Heller, Alan McNaught, Igor Pletnev, Stephen Stein, and Dmitrii Tchekhovskoi. 2015. InChI, the IUPAC International Chemical Identifier. Journal of Cheminformatics 7, 1 (2015), 23. [43]Ask Hjorth Larsen, Jens JĂžrgen Mortensen, Jakob Blomqvist, Ivano E Castelli, Rune Christensen, Marcin DuĆak, Jesper Friis, Michael N Groves, BjĂžrk Hammer, Cory Hargus, et al.2017. The atomic simulation environmentâa Python library for working with atoms. Journal of Physics: Condensed Matter 29, 27 (2017), 273002. [44]Hisayuki Horai, Masanori Arita, Shigehiko Kanaya, Yoshito Nihei, Tasuku Ikeda, Kazuhiro Suwa, Yuya Ojima, Kenichi Tanaka, Satoshi Tanaka, Ken Aoshima, et al.2010. MassBank: a public repository for sharing mass spectral data for life sciences. Journal of mass spectrometry 45, 7 (2010), 703â714. [45]Xiuyuan Hu, Guoqing Liu, Yang Zhao, and Hao Zhang. 2023. De novo drug design using reinforcement learning with multiple gpt agents. Advances in Neural Information Processing Systems 36 (2023), 7405â7418. [46]Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Connor W. Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik. 2021. Therapeutics Data Commons: machine learning datasets and tasks for drug discovery and development. arXiv preprint arXiv:2102.09548 (2021). [47]Linda Hung, Joyce A. Yager, Danielle Monteverde, Dave Baiocchi, Ha-Kyung Kwon, Shijing Sun, and Santosh Suram. 2024. Autonomous laboratories for accelerated materials discovery: A community survey and practical insights. Digital Discovery 3 (2024), 1273â1279. doi:10.1039/D4D00059E [48]Yoshitaka Inoue, Tianci Song, Xinling Wang, Augustin Luna, and Tianfan Fu. 2025. Drugagent: Multi-agent large language model-based reasoning for drug- target interaction prediction. ArXiv (2025), arXivâ2408. [49]John J Irwin, Khanh G Tang, Jennifer Young, Chinzorig Dandarchuluun, Ben- jamin R Wong, Munkhzul Khurelbaatar, Yurii S Moroz, John Mayfield, and Roger A Sayle. 2020. ZINC20âa free ultralarge-scale chemical database for ligand discovery. Journal of chemical information and modeling 60, 12 (2020), 6065â6073. [50] Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, et al.2013. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL materials 1, 1 (2013). [51] Wengong Jin, Regina Barzilay, and Tommi Jaakkola. 2018. Junction tree varia- tional autoencoder for molecular graph generation. In International conference on machine learning. PMLR, 2323â2332. [52]Y Joly. 2001. X-ray absorption near-edge structure calculations beyond the muffin-tin approximation. Physical Review B 63, 12 (2001), 125120. [53]Yeonghun Kang and Jihan Kim. 2024. ChatMOF: an artificial intelligence system for predicting and generating metal-organic frameworks using large language models. Nature communications 15, 1 (2024), 4705. [54]Hyomin Kim, Yunhui Jang, and Sungsoo Ahn. 2025. Mt-mol: Multi agent system with tool-based reasoning for molecular optimization. Artificial Intelligence Repository (2025). [55] Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, et al.2023. Pub- Chem 2023 update. Nucleic acids research 51, D1 (2023), D1373âD1380. [56]Takayuki Kimura. 2026. VQ-Atom: Semantic discretization of local atomic envi- ronments for molecular representation learning. arXiv preprint arXiv:2605.16823 (2026). [57]Stephen T. Knox, Sam J. Parkinson, Clarissa Y. P. Wilding, Richard A. Bourne, and Nicholas J. Warren. 2022. Autonomous polymer synthesis delivered by multi-objective closed-loop optimisation. Polymer Chemistry 13, 11 (2022), 1576â1585. doi:10.1039/D2PY00040G [58]Mario Krenn, Florian Hase, AkshatKumar Nigam, Pascal Friederich, and AlĂĄn Aspuru-Guzik. 2020. SELFIES: a robust representation of semantically con- strained graphs with an example application in chemistry. Machine Learning: Science and Technology 1, 4 (2020), 045024. [59] Stefan Kuhn, Heinz Kolshorn, Christoph Steinbeck, and Nils Schlörer. 2024. Twenty years of nmrshiftdb2: A case study of an open database for analytical chemistry. Magnetic Resonance in Chemistry 62, 2 (2024), 74â83. [60] Vinay Kumar, Satyendra Rajput, Mausam, and N. M. Anoop Krishnan. 2026. MDGym: Benchmarking AI Agents on Molecular Simulations. arXiv preprint arXiv:2605.08941 (2026). arXiv:2605.08941 [cs.AI] https://arxiv.org/abs/2605. 08941 [61]Khiem Le, Ting Hua, and Nitesh V Chawla. 2024. AgentDrug: Utilizing Large Language Models in An Agentic Workflow for Zero-Shot Molecular Optimiza- tion. arXiv preprint arXiv:2410.13147 (2024). [62]Namkyeong Lee, Edward De Brouwer, Ehsan Hajiramezanali, Tommaso Bian- calani, Chanyoung Park, and Gabriele Scalia. 2026. Rag-enhanced collaborative llm agents for drug discovery. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 561â569. [63]Shi Xuan Leong, Caleb E. Griesbach, Rui Zhang, Kourosh Darvish, Yuchi Zhao, Abhijoy Mandal, Yunheng Zou, Han Hao, Varinia Bernales, AlĂĄn Aspuru-Guzik, et al.2025. Steering towards safe self-driving laboratories. Nature Reviews Chemistry 9 (2025), 707â722. doi:10.1038/s41570-025-00747-x [64] Hao Li, He Cao, Shenyao Peng, Zijing Liu, Bin Feng, Yu Wang, Zhiyuan Yan, Yonghong Tian, Yu Li, and Li Yuan. 2026. Agentic reinforcement learning empowers next-generation chemical language models for molecular design and synthesis. arXiv preprint arXiv:2601.17687 (2026). [65] Jiatong Li, Junxian Li, Weida Wang, Yunqing Liu, Changmeng Zheng, Yatao Bian, Dongzhan Zhou, Xiao-Yong Wei, and Qing Li. 2026. Speak-to-structure: Evaluating llms in open-domain natural language-driven molecule generation. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 9314â9325. PreprintConferenceâ17, July 2017, Washington, DC, USA [66]Jiatong Li, Wei Liu, Zhihao Ding, Wenqi Fan, Yuqiang Li, and Qing Li. 2025. Large language models are in-context molecule learners. IEEE Transactions on Knowledge and Data Engineering (2025). [67] Jiatong Li, Yunqing Liu, Wenqi Fan, Xiao-Yong Wei, Hui Liu, Jiliang Tang, and Qing Li. 2024. Empowering molecule discovery for molecule-caption transla- tion with large language models: A chatgpt perspective. IEEE transactions on knowledge and data engineering 36, 11 (2024), 6071â6083. [68]Jiatong Li, Yunqing Liu, Wei Liu, Jingdi Lei, Di Zhang, Wenqi Fan, Dongzhan Zhou, Yuqiang Li, and Qing Li. 2026. Molreflect: Towards in-context fine-grained alignments between molecules and texts. IEEE Transactions on Knowledge and Data Engineering (2026). [69]Jiatong Li, Yuxuan Ren, Weida Wang, Xiaoyong Wei, and Yatao Bian. 2026. Chemical Chain-of-Thought Functions as a Hallucination-Prone Molecular Scratchpad. arXiv preprint arXiv:2607.20935 (2026). [70]Jiatong Li, Yuxuan Ren, Weida Wang, Changmeng Zheng, Xiao-yong Wei, Qing Li, and Yatao Bian. 2026. MolViBench: Evaluating LLMs on Molecular Vibe Coding. arXiv preprint arXiv:2605.02351 (2026). [71] Jiatong Li, Weida Wang, Qinggang Zhang, Junxian Li, Di Zhang, Changmeng Zheng, Shufei Zhang, Xiaoyong Wei, and Qing Li. 2025. Mol-r1: Towards explicit long-cot reasoning in molecule discovery. arXiv preprint arXiv:2508.08401 (2025). [72]Jiatong Li, Weida Wang, Changmeng Zheng, Shufei Zhang, Yatao Bian, Xiao yong Wei, and Qing Li. 2026. Do LLMs Truly Generalize in the Molecular Domain? A Perturbation-Based Analysis. arXiv preprint arXiv:2607.01800 (2026). [73]Junxian Li, Di Zhang, Xunzhi Wang, Zeying Hao, Jingdi Lei, Qian Tan, Cai Zhou, Wei Liu, Yaotian Yang, Xinrui Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Wei Li, Shufei Zhang, Mao Su, Wanli Ouyang, Yuqiang Li, and Dongzhan Zhou. 2025. ChemVLM: Exploring the Power of Multimodal Large Language Models in Chemistry Area. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 415â423. doi:10.1609/aaai.v39i1.32020 [74] Kun Li, Zhennan Wu, Shoupeng Wang, Jia Wu, Shirui Pan, and Wenbin Hu. 2025. DrugPilot: LLM-based parameterized reasoning agent for drug discovery. arXiv preprint arXiv:2505.13940 (2025). [75]Lingxiao Li, Haobo Zhang, Ruohao Fan, Bin Chen, and Jiayu Zhou. 2026. Molecu- lar Lead Optimization via Agentic Tool Planning. arXiv preprint arXiv:2605.28862 (2026). [76]Yizhan Li, Florence Cloutier, Sifan Wu, Ali Parviz, Boris Knyazev, Yan Zhang, Glen Berseth, and Bang Liu. 2026. MË 4olGen: Multi-Agent, Multi-Stage Molec- ular Generation under Precise Multi-Property Constraints. arXiv preprint arXiv:2601.10131 (2026). [77]Zhucong Li, Jin Xiao, Bowei Zhang, Zhijian Zhou, Qianyu He, Fenglei Cao, Jiaqing Liang, and Yuan Qi. 2025. Chemhts: Hierarchical tool stacking for enhancing chemical agents. arXiv preprint arXiv:2502.14327 (2025). [78]Zhucong Li, Bowei Zhang, Jin Xiao, Zhijian Zhou, Fenglei Cao, Jiaqing Liang, and Yuan Qi. 2025. Chemhas: Hierarchical agent stacking for enhancing chemistry tools. arXiv preprint arXiv:2505.21569v1 (2025). [79] P Linstorm. 1998. NIST chemistry webbook, NIST standard reference database number 69. J. Phys. Chem. Ref. Data, Monograph 9 (1998), 1â1951. [80]Gang Liu, Michael Sun, Wojciech Matusik, Meng Jiang, and Jie Chen. 2025. Mul- timodal large language models for inverse molecular design with retrosynthetic planning. In International Conference on Learning Representations, Vol. 2025. 41744â41771. [81]Siyu Liu, Bo Hu, Beilin Ye, Jiamin Xu, David J Srolovitz, and Tongqi Wen. 2025. Mattools: Benchmarking large language models for materials science tools. arXiv preprint arXiv:2505.10852 (2025). [82]Sizhe Liu, Yizhou Lu, Siyu Chen, Xiyang Hu, Jieyu Zhao, Yingzhou Lu, and Yue Zhao. 2024. Drugagent: Automating ai-aided drug discovery programming through llm multi-agent collaboration. arXiv preprint arXiv:2411.15692 (2024). [83]Shengchao Liu, Hanchen Wang, Weiyang Liu, Joan Lasenby, Hongyu Guo, and Jian Tang. 2022. Pre-training molecular graph representation with 3D geometry. In International Conference on Learning Representations. [84]Shengchao Liu, Jiongxiao Wang, Yijin Yang, Chengpeng Wang, Ling Liu, Hongyu Guo, and Chaowei Xiao. 2024. Conversational Drug Editing Using Retrieval and Domain Feedback. In The Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=yRrPfKyJQ2 [85]Tianyu Liu, Allen Xin Wang, Antonia Panescu, Lisa Xinyi Chen, Wenxin Long, Xinyu Wei, et al.2026. Benchmarking AI Agents for Addressing Scientific Challenges Across Scales. arXiv preprint arXiv:2606.12736 (2026). https://arxiv. org/abs/2606.12736 [86] Wei Liu, Jiangtao Feng, Hongli Yu, Yuxuan Song, Yuqiang Li, Shufei Zhang, Lei Bai, Wei-Ying Ma, and Hao Zhou. 2026. Retro-r1: LLM-based agentic retrosynthe- sis. Advances in Neural Information Processing Systems 38 (2026), 70709â70737. [87]Xuhan Liu, Kexin Ye, Herman W. T. van Vlijmen, Adriaan P. IJzerman, and Gerard J. P. van Westen. 2021. DrugEx v2: De novo design of drug molecules by Pareto-based multi-objective reinforcement learning in polypharmacology. Journal of Cheminformatics 13 (2021), 85. doi:10.1186/s13321-021-00561-9 [88]Hannes H. Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Artur Voronov, Lewis H. Mervin, Ola Engkvist, and Hongming Chen. 2024. REINVENT4: Modern AI-driven generative molecule design. Journal of Cheminformatics 16 (2024), 20. doi:10.1186/s13321-024-00812-5 [89]Daniel M Lowe, Peter T Corbett, Peter Murray-Rust, and Robert C Glen. 2011. Chemical name to structure: OPSIN, an open source solution. [90] Xufang Luo, Yuge Zhang, Zhiyuan He, Zilong Wang, Siyun Zhao, Dongsheng Li, Luna K. Qiu, and Yuqing Yang. 2025. Agent Lightning: Train ANY AI Agents with Reinforcement Learning. arXiv preprint arXiv:2508.03680 (2025). [91] Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2024. Augmenting large language models with chemistry tools. Nature machine intelligence 6, 5 (2024), 525â535. doi:10.1038/s42256-024- 00832-8 [92]Shreshth A. Malik, Tiarnan Doherty, Panagiotis Tigas, Muhammed Razzak, Stephen J. Roberts, Aron Walsh, and Yarin Gal. 2026. MADE: Benchmark Environments for Closed-Loop Materials Discovery. In Proceedings of the 43rd International Conference on Machine Learning, Vol. 306. https://arxiv.org/abs/ 2601.20996 [93]Andrew D McNaughton, Gautham Krishna Sankar Ramalaxmi, Agustin Kruel, Carter R Knutson, Rohith A Varikoti, and Neeraj Kumar. 2024. Cactus: Chemistry agent connecting tool usage to science. ACS omega 9, 46 (2024), 46563â46573. doi:10.1021/acsomega.4c08408 [94]Andrew T McNutt, Paul Francoeur, Rishal Aggarwal, Tomohide Masuda, Rocco Meli, Matthew Ragoza, Jocelyn Sunseri, and David Ryan Koes. 2021. GNINA 1.0: molecular docking with deep learning. Journal of cheminformatics 13, 1 (2021), 43. [95]Siddharth Narayanan, James Braza, Ryan-Rhys Griffiths, Albert Bou, Geemi Wellawatte, Mayk Caldas Ramos, Ludovico Mitchener, Michael Pieler, Sam Rodriques, and Andrew White. 2026. Training a scientific reasoning model for chemistry. Advances in Neural Information Processing Systems 38 (2026), 157671â157710. [96]Frank Neese, Frank Wennmohs, Ute Becker, and Christoph Riplinger. 2020. The ORCA quantum chemistry program package. The Journal of chemical physics 152, 22 (2020). [97] Steffen Neumann, RenĂ© Meier, Michael Wenk, Anjana Elapavalore, Takaaki Nishioka, Tobias Schulze, Michael Stravs, Hiroshi Tsugawa, Fumio Matsuda, and Emma L Schymanski. 2026. MassBank: an open and FAIR mass spectral data resource. Nucleic Acids Research 54, D1 (2026), D601âD606. [98]Thao Nguyen and Heng Ji. 2026. MolLingo: Molecule-Native Representations for LLM-Powered Scientific Agents. arXiv preprint arXiv:2605.27853 (2026). https://arxiv.org/abs/2605.27853 [99] Liangbo Ning, Ziran Liang, Zhuohang Jiang, Haohao Qu, Yujuan Ding, Wenqi Fan, Xiao-yong Wei, Shanru Lin, Hui Liu, Philip S Yu, et al.2025. A survey of webagents: Towards next-generation ai agents for web automation with large foundation models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 6140â6150. [100] Emmanuel Noutahi, Cristian Gabellini, Michael Craig, Jonathan S. C. Lim, and Prudencio Tossou. 2023. Gotta be SAFE: a new framework for molecular design. arXiv preprint arXiv:2310.10773 (2023). [101]Noel M OâBoyle, Michael Banck, Craig A James, Chris Morley, Tim Vandermeer- sch, and Geoffrey R Hutchison. 2011. Open Babel: An open chemical toolbox. Journal of cheminformatics 3, 1 (2011), 33. [102]Janghoon Ock, Radheesh Sharma Meda, Srivathsan Badrinarayanan, Neha S Aluru, Achuth Chandrasekhar, and Amir Barati Farimani. 2026. Large language model agent for modular task execution in drug discovery. Journal of Chemical Information and Modeling 66, 4 (2026), 2055â2068. [103] Marcus Olivecrona, Thomas Blaschke, Ola Engkvist, and Hongming Chen. 2017. Molecular de-novo design through deep reinforcement learning. Journal of Cheminformatics 9, 1 (2017), 48. doi:10.1186/s13321-017-0235-x [104]Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L Chevrier, Kristin A Persson, and Gerbrand Ceder. 2013. Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis. Computational Materials Science 68 (2013), 314â319. [105]Qihua Pan, Dong Xu, Qianwei Yang, Jenna Xinyi Yao, Sisi Yuan, Zexuan Zhu, Jianqiang Li, and Junkai Ji. 2026. FROGENT: An End-to-End Full-process Drug Design Multi-Agent System. arXiv:2508.10760 [q-bio.BM] https://arxiv.org/abs/ 2508.10760 [106]Gihan Panapitiya, Emily Saldanha, Heather Job, and Olivia Hess. 2026. Autolabs: Cognitive multi-agent systems with self-correction for autonomous chemical experimentation. Scientific Reports 16, 1 (2026), 19554. doi:10.1038/s41598-026- 45593-z [107] Robert M Parrish, Lori A Burns, Daniel GA Smith, Andrew C Simmonett, A Eu- gene DePrince I, Edward G Hohenstein, Ugur Bozkaya, Alexander Yu Sokolov, Roberto Di Remigio, Ryan M Richard, et al.2017. Psi4 1.1: An open-source electronic structure program emphasizing automation, advanced libraries, and interoperability. Journal of chemical theory and computation 13, 7 (2017), 3185â 3197. [108]Iman Peivaste, Ahmed Makradi, and Salim Belouettar. 2026. ChemNavigator: Agentic AI Discovery of Design Rules for Organic Photocatalysts. arXiv preprint arXiv:2601.17084 (2026). PreprintMolLLMAgent [109]Thang D Pham, Aditya Tanikanti, and Murat Keçeli. 2026. ChemGraph as an agentic framework for computational chemistry workflows. Communications Chemistry (2026). [110] Daniil Polykovskiy, Alexander Zhebrak, Benjamin Sanchez-Lengeling, Sergey Golovanov, Oktai Tatanov, Stanislav Belyaev, Rauf Kurbanov, Aleksey Arta- monov, Vladimir Aladinskiy, Mark Veselov, et al.2020. Molecular sets (MOSES): a benchmarking platform for molecular generation models. Frontiers in Phar- macology 11 (2020), 565644. [111] Mariya Popova, Olexandr Isayev, and Alexander Tropsha. 2018. Deep reinforce- ment learning for de novo drug design. Science Advances 4, 7 (2018), eaap7885. doi:10.1126/sciadv.aap7885 [112]Yujie Qian, Jiang Guo, Zhengkai Tu, Connor W. Coley, and Regina Barzilay. 2023. MolScribe: Robust molecular structure recognition with image-to-graph generation. In International Conference on Learning Representations. [113]Nian Ran, Yue Wang, Xiaoyuan Zhang, Zhongzheng Li, Qingsong Ran, Wenhao Li, and Richard Allmendinger. 2025. ExLLM: Experience-Enhanced LLM Op- timization for Molecular Design and Beyond. arXiv preprint arXiv:2502.12845 (2025). [114]Jerret Ross, Brian Belgodere, Vijil Chenthamarakshan, Inkit Padhi, Youssef Mroueh, and Payel Das. 2022. Large-scale chemical language representations capture molecular structure and properties. Nature Machine Intelligence 4, 12 (2022), 1256â1264. [115]Samuel Rothfarb, Megan C Davis, Ivana Matanovic, Baikun Li, Edward F Holby, and Wilton JM Kort-Kamp. 2025. Hierarchical Multi-agent Large Language Model Reasoning for Autonomous Functional Materials Discovery. arXiv preprint arXiv:2512.13930 (2025). [116]Yixiang Ruan, Chenyin Lu, Ning Xu, Yuchen He, Yixin Chen, Jian Zhang, Jun Xuan, Jianzhang Pan, Qun Fang, Hanyu Gao, et al.2024. An automatic end- to-end chemical synthesis development platform powered by large language models. Nature communications 15, 1 (2024), 10160. doi:10.1038/s41467-024- 54457-x [117] Nicholas T. Runcie, Fergus Imrie, and Charlotte M. Deane. 2026. Molecular representations for large language models. arXiv preprint arXiv:2605.01822 (2026). https://arxiv.org/abs/2605.01822 [118] Timo Schick, Jane Dwivedi-Yu, Roberto DessĂŹ, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. In Advances in Neural Information Processing Systems, Vol. 36. 68539â68551. [119] Kristof T. SchĂŒt, Pieter-Jan Kindermans, Huziel E. Sauceda, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert MĂŒller. 2017. SchNet: A continuous- filter convolutional neural network for modeling quantum interactions. In Advances in Neural Information Processing Systems, Vol. 30. [120]Yujiong Shen, Yajie Yang, Zhiheng Xi, Binze Hu, Huayu Sha, Jiazheng Zhang, Qiyuan Peng, Junlin Shang, Jixuan Huang, Yutao Fan, Jingqi Tong, Shihan Dou, Ming Zhang, Lei Bai, Zhenfei Yin, Tao Gui, Xingjun Ma, Qi Zhang, Xu- anjing Huang, and Yu-Gang Jiang. 2026. SciAgentGym: Benchmarking Multi- Step Scientific Tool-use in LLM Agents. arXiv preprint arXiv:2602.12984 (2026). arXiv:2602.12984 [cs.CL] https://arxiv.org/abs/2602.12984 [121] Zhuofan Shi, Yufei Shao, Mengyan Dai, Yadong Yu, Pan Xiang, Dongliang Huang, Hongxu An, Chunxiao Xin, Haiyang Shen, Zhenyu Wang, et al.2026. MDAgent2: Large Language Model for Code Generation and Knowledge Q&A in Molecular Dynamics. arXiv preprint arXiv:2601.02075 (2026). [122]Benjamin J. Shields, Jason Stevens, Jun Li, Marvin Parasram, Farhan Damani, Jesus I. M. Alvarado, Jacob M. Janey, Ryan P. Adams, and Abigail G. Doyle. 2021. Bayesian reaction optimization as a tool for chemical synthesis. Nature 590 (2021), 89â96. doi:10.1038/s41586-021-03213-y [123]Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366 (2023). [124]Zachary S. Siegel, Sayash Kapoor, Nitya Nadgir, Benedikt Stroebl, and Arvind Narayanan. 2024. CORE-Bench: Fostering the Credibility of Published Research Through a Computational Reproducibility Agent Benchmark. arXiv preprint arXiv:2409.11363 (2024). https://arxiv.org/abs/2409.11363 [125]Gleb V Solovev, Alina B Zhidkovskaya, Anastasia Orlova, Nina Gubina, Anas- tasia Vepreva, Rodion Golovinskii, Ilya Tonkii, Ivan Dubrovsky, Ivan Gurev, Dmitry Gilemkhanov, et al.2025. MADD: Multi-Agent Drug Discovery Orches- tra. In Findings of the Association for Computational Linguistics: EMNLP 2025. 6956â6998. [126] Tao Song, Man Luo, Xiaolong Zhang, Linjiang Chen, Yan Huang, Jiaqi Cao, Qing Zhu, Daobin Liu, Baicheng Zhang, Gang Zou, et al.2025. A multiagent-driven robotic AI chemist enabling autonomous chemical research on demand. Journal of the American Chemical Society 147, 15 (2025), 12534â12545. doi:10.1021/jacs. 4c17738 [127]Henry Sprueill, Carl Edwards, Mariefel Olarte, Udishnu Sanyal, Heng Ji, and Sutanay Choudhury. 2023. Monte carlo thought search: Large language model querying for complex scientific reasoning in catalyst design. In Findings of the Association for Computational Linguistics: EMNLP 2023. 8348â8365. [128]Henry W Sprueill, Carl Edwards, Khushbu Agarwal, Mariefel V Olarte, Udishnu Sanyal, Conrad Johnston, Hongbin Liu, Heng Ji, and Sutanay Choudhury. 2024. ChemReasoner: Heuristic search over a large language modelâs knowledge space using quantum-chemical feedback. arXiv preprint arXiv:2402.10980 (2024). [129]Christoph Steinbeck, Stefan Krause, and Stefan Kuhn. 2003. NMRShiftDB constructing a free chemical information system with open-source components. Journal of chemical information and computer sciences 43, 6 (2003), 1733â1739. [130]Minyi Su, Qifan Yang, Yu Du, Guoqin Feng, Zhihai Liu, Yan Li, and Renxiao Wang. 2018. Comparative assessment of scoring functions: the CASF-2016 update. Journal of chemical information and modeling 59, 2 (2018), 895â913. [131]Nathan J. Szymanski, B. Rendy, Y. Fei, R. E. Kumar, T. He, D. Milsted, M. J. McDermott, M. Gallant, E. D. Cubuk, A. Merchant, et al.2023. An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624 (2023), 86â91. doi:10.1038/s41586-023-06734-w [132]Elham Tabassi. 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). Technical Report NIST AI 100-1. National Institute of Standards and Technology. doi:10.6028/NIST.AI.100-1 [133] Xiangru Tang, Tianyu Hu, Muyang Ye, Yanjun Shao, Xunjian Yin, Siru Ouyang, Wangchunshu Zhou, Pan Lu, Zhuosheng Zhang, Yilun Zhao, et al.2025. Chema- gent: Self-updating memories in large language models improves chemical reasoning. In The Thirteenth International Conference on Learning Representa- tions. [134]Xiangru Tang, Qiao Jin, Kunlun Zhu, Tongxin Yuan, Yichi Zhang, Wangchunshu Zhou, Meng Qu, Yilun Zhao, Jian Tang, Zhuosheng Zhang, Arman Cohan, Zhiyong Lu, and Mark B. Gerstein. 2024. Risks of AI scientists: prioritizing safeguarding over autonomy. Nature Communications 16 (2024). doi:10.1038/ s41467-025-63913-1 [135] Asu BĂŒĆra Temizer, Gökçe UludoÄan, Rıza Ăzçelik, Taha Koulani, Elif Ozkirimli, Kutlu O. Ulgen, NilgĂŒn Karalı, and Arzucan ĂzgĂŒr. 2022. Exploring data-driven chemical SMILES tokenization approaches to identify key protein-ligand binding moieties. arXiv preprint arXiv:2210.14642 (2022). [136] Morgan Thomas et al.2022. Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation. Journal of Cheminformatics 14 (2022), 68. doi:10.1186/s13321-022-00646-z [137] Aidan P Thompson, H Metin Aktulga, Richard Berger, Dan S Bolintineanu, W Michael Brown, Paul S Crozier, Pieter J Inât Veld, Axel Kohlmeyer, Stan G Moore, Trung Dac Nguyen, et al.2022. LAMMPS-a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer physics communications 271 (2022), 108171. [138]Minyang Tian, Luyu Gao, Shizhuo Dylan Zhang, Xinan Chen, Cunwei Fan, Xuefei Guo, Roland Haas, Pan Ji, Kittithat Krongchon, Yao Li, Shengyan Liu, Di Luo, Yutao Ma, Hao Tong, Kha Trinh, Chenyu Tian, Zihan Wang, Bohao Wu, Yanyu Xiong, Shengzhu Yin, Minhui Zhu, Kilian Lieret, Yanxin Lu, Genglin Liu, Yufeng Du, Tianhua Tao, Ofir Press, Jamie Callan, Eliu Huerta, and Hao Peng. 2024. SciCode: A Research Coding Benchmark Curated by Scientists. In Advances in Neural Information Processing Systems, Vol. 37. doi:10.52202/079017-0963 Datasets and Benchmarks Track. [139]Benjamin I Tingle, Khanh G Tang, Mar Castanon, John J Gutierrez, Munkhzul Khurelbaatar, Chinzorig Dandarchuluun, Yurii S Moroz, and John J Irwin. 2023. ZINC-22: A free multi-billion-scale database of tangible compounds for ligand discovery. Journal of chemical information and modeling 63, 4 (2023), 1166â1176. [140]Oleg Trott and Arthur J Olson. 2010. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of computational chemistry 31, 2 (2010), 455â461. [141]Zhengkai Tu, Sourabh J Choure, Mun Hong Fong, Jihye Roh, Itai Levin, Kevin Yu, Joonyoung F Joung, Nathan Morgan, Shih-Cheng Li, Xiaoqi Sun, et al.2025. ASKCOS: an open source software suite for synthesis planning. arXiv preprint arXiv:2501.01835 (2025). [142]UniProt Consortium. 2023. UniProt: the universal protein knowledgebase in 2023. Nucleic acids research 51, D1 (2023), D523âD531. [143]Fabio Urbina, Filippa Lentzos, CĂ©dric Invernizzi, and Sean Ekins. 2022. Dual use of artificial-intelligence-powered drug discovery. Nature Machine Intelligence 4 (2022), 189â191. doi:10.1038/s42256-022-00465-9 [144]Ivana Vichentijevikj, Kostadin Mishev, and Monika Simjanoska Misheva. 2026. Prompt-to-pill: Multi-Agent drug discovery and clinical simulation pipeline. Bioinformatics Advances 6, 1 (2026), vbaf323. [145]Amanda A. Volk and Milad Abolhasani. 2024. Performance metrics to unleash the power of self-driving labs in chemistry and materials science. Nature Communications 15 (2024), 1378. doi:10.1038/s41467-024-45569-5 [146] Amanda A. Volk, Robert W. Epps, Daniel T. Yonemoto, Benjamin S. Masters, Felix Castellano, Kristofer G. Reyes, and Milad Abolhasani. 2023. AlphaFlow: Autonomous discovery and optimization of multi-step chemistry using a self- driven fluidic lab guided by reinforcement learning. Nature Communications 14, 1 (2023), 1403. doi:10.1038/s41467-023-37139-y [147]Alexius Wadell, Anoushka Bhutani, and Venkatasubramanian Viswanathan. 2024.Tokenization for molecular foundation models. arXiv preprint arXiv:2409.15370 (2024). PreprintConferenceâ17, July 2017, Washington, DC, USA [148]Renxiao Wang, Xueliang Fang, Yipin Lu, Chao-Yie Yang, and Shaomeng Wang. 2005. The PDBbind database: methodologies and updates. Journal of medicinal chemistry 48, 12 (2005), 4111â4119. [149] Weida Wang, Benteng Chen, Di Zhang, Wanhao Liu, Ben Gao, Shuchen Pu, Shuzhou Sun, Jin Zeng, Tianshu Yu, Wanli Ouyang, et al.2026. Chem-r: Learning to reason as a chemist. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 12291â12302. [150]Ziqing Wang and Kaize Ding. 2025. ReMol: LLM-guided Molecular Optimization with Reinforcement Learning. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (PhD Consortium). ACM. [151]David Weininger. 1988. SMILES, a chemical language and information sys- tem. 1. Introduction to methodology and encoding rules. Journal of Chemical Information and Computer Sciences 28, 1 (1988), 31â36. [152]Mengsong Wu, YaFei Wang, Yidong Ming, Yuqi An, Yuwei Wan, Wenliang Chen, Binbin Lin, Yuqiang Li, Tong Xie, and Dongzhan Zhou. 2025. Chematagent: Enhancing llms for chemistry and materials science through tree-search based tool learning. arXiv preprint arXiv:2506.07551 (2025). [153] Yuyang Wu, Yue Huang, Shuaike Shen, Xujian Wang, Shuhao Zhang, Qiyao Xue, Weichen Liu, Runtian Gao, Jian Ma, Xiangliang Zhang, and Olexandr Isayev. 2026. Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning. arXiv preprint arXiv:2605.07251 (2026). arXiv:2605.07251 https: //arxiv.org/abs/2605.07251 [154]Zhenqin Wu, Bharath Ramsundar, Evan N. Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S. Pappu, Karl Leswing, and Vijay Pande. 2018. MoleculeNet: a benchmark for molecular machine learning. Chemical Science 9, 2 (2018), 513â530. [155]Yutong Xie, Chence Shi, Hao Zhou, Yuwei Yang, Weinan Zhang, Yong Yu, and Lei Li. 2021. MARS: Markov molecular sampling for multi-objective drug discovery. In International Conference on Learning Representations. [156] Qiang Xu, Shengyuan Bai, Leqing Chen, Zijing Liu, and Yu Li. 2025. ChemLabs on ChemO: A Multi-Agent System for Multimodal Reasoning on IChO 2025. arXiv preprint arXiv:2511.16205 (2025). [157]Zhiyuan Yan, Chen Liu, Boxuan Zhao, Kaiqing Lin, Jixiang Zhao, Yimi Wang, Liuzhenghao Lv, Hao Li, Shanzhuo Zhang, Li Yuan, et al.2026. MoleCode unlocks structural intelligence in large language models. arXiv preprint arXiv:2605.16480 (2026). https://arxiv.org/abs/2605.16480 [158] Kevin Yang, Kyle Swanson, Wengong Jin, Connor Coley, Philipp Eiden, Hua Gao, Angel Guzman-Perez, Timothy Hopper, Brian Kelley, Miriam Mathea, et al. 2019. Analyzing learned molecular representations for property prediction. Journal of Chemical Information and Modeling 59, 8 (2019), 3370â3388. [159] Yu Yang et al.2024. Mol-AIR: Molecular reinforcement learning with adaptive intrinsic rewards. Journal of Chemical Information and Modeling (2024). doi:10. 1021/acs.jcim.4c01669 [160] Zaifei Yang, Weiyu Chen, Yaqing Wang, and James Kwok. 2026. Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure- Based Drug Design. arXiv preprint arXiv:2606.00555 (2026). [161]Zerui Yang, Yuwei Wan, Siyu Yan, Yudai Matsuda, Tong Xie, Bram Hoex, and Linqi Song. 2025. DrugMCTS: A drug repurposing framework combining multi- agent, RAG and Monte Carlo Tree Search. arXiv preprint arXiv:2507.07426 (2025). [162] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In International Conference on Learning Representations. [163] Geyan Ye, Xibao Cai, Houtim Lai, Xing Wang, Junhong Huang, Longyue Wang, Wei Liu, and Xiangxiang Zeng. 2025. Drugassist: A large language model for molecule optimization. Briefings in Bioinformatics 26, 1 (2025), bbae693. doi:10.1093/bib/bbae693 [164]Xinwu Ye, Yicheng Mao, Jia Zhang, Yimeng Liu, Li Hao, Fang Wu, Zhiwei Li, Yuxuan Liao, Zehong Wang, Yingcheng Wu, et al.2026. Latentchem: From tex- tual cot to latent thinking in chemical reasoning. arXiv preprint arXiv:2602.07075 (2026). [165]Jiaxuan You, Bowen Liu, Rex Ying, Vijay Pande, and Jure Leskovec. 2018. Graph convolutional policy network for goal-directed molecular graph generation. arXiv preprint arXiv:1806.02473 (2018). [166]Botao Yu, Frazier N Baker, Ziru Chen, Garrett Herb, Boyu Gou, Daniel Adu- Ampratwum, Xia Ning, and Huan Sun. 2025. Tooling or not tooling? the impact of tools on language agents for chemistry problem solving. In Findings of the Association for Computational Linguistics: NAACL 2025. 7620â7640. [167] Jiajun Yu, Yizhen Zheng, Huan Yee Koh, Shirui Pan, Tianyue Wang, and Haishuai Wang. 2025. Collaborative expert llms guided multi-objective molecular opti- mization. arXiv preprint arXiv:2503.03503 (2025). [168]Jie Yue, Bingxin Peng, Yu Chen, Jieyu Jin, Xinda Zhao, Chao Shen, Xiangyang Ji, Chang-Yu Hsieh, Jianfei Song, Tingjun Hou, et al.2024. Unlocking compre- hensive molecular design across all scenarios with large language model and unordered chemical language. Chemical Science 15, 34 (2024), 13727â13740. [169]Atakan YĂŒksel, Erva Ulusoy, Atabey ĂnlĂŒ, and Tunca DoÄan. 2023. SELFormer: Molecular representation learning via SELFIES language models. arXiv preprint arXiv:2304.04662 (2023). [170]Barbara Zdrazil, Eloy Felix, Fiona Hunter, Emma J Manners, James Blackshaw, Sybilla Corbett, Marleen De Veij, Harris Ioannidis, David Mendez Lopez, Juan F Mosquera, et al.2024. The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic acids research 52, D1 (2024), D1180âD1192. [171]Huan Zhang, Yizhan Li, Wenhao Huang, Ziyu Hou, Yu Song, Xuye Liu, Farshid Effaty, Jinya Jiang, Sifan Wu, Qianggang Ding, et al.2026. Towards Agentic Intelligence for Materials Science. arXiv preprint arXiv:2602.00169 (2026). [172] Jinwei Zhang, Xucheng Liang, Yu Zhang, Ruijie Yu, Xiaokang Yang, Yaohui Jin, and Yanyan Xu. 2026. ChemReason-Bench: Benchmarking Large Language Models for Procedural Reasoning in Experimental Chemistry. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 33211â33248. doi:10.18653/v1/2026.acl-long.1535 [173]Jia Zhang, Tengfei Ma, Tianle Li, Daojian Zeng, Xieping Gao, and Xiangxiang Zeng. 2026. Agents on a Tree: Pathwise Coordination for Multi-Objective Molecular Optimization. arXiv preprint arXiv:2606.00008 (2026). [174] Lisheng Zhang, Lilong Wang, Xiangyu Sun, Wei Tang, Haoyang Su, Yuehui Qian, Qikui Yang, Qingsong Li, Zhenyu Tang, Haoran Sun, Yingnan Han, Yankai Jiang, Wenjie Lou, Bowen Zhou, Xiaosong Wang, Lei Bai, and Zhengwei Xie. 2026. MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization. arXiv preprint arXiv:2604.21937 (2026). https://arxiv.org/abs/2604.21937 [175]Mingxu Zhang, Dazhong Shen, and Ying Sun. 2025. AtomDisc: An atom-level to- kenizer that boosts molecular LLMs and reveals structure-property associations. arXiv preprint arXiv:2512.03080 (2025). [176]Wengyu Zhang, Xiao-Yong Wei, and Qing Li. 2026. Mol-Debate: multi-agent debate improves structural reasoning in molecular design. arXiv preprint arXiv:2604.20254 (2026). [177] Yu Zhang, Ruijie Yu, Jidong Tian, Feng Zhu, Jiapeng Liu, Xiaokang Yang, Yaohui Jin, and Yanyan Xu. 2025. ChemActor: Enhancing Automated Extraction of Chemical Synthesis Actions with LLM-Generated Data. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 24291â24314. [178] Changmeng Zheng, Dayong Liang, Wengyu Zhang, Xiao-Yong Wei, Tat-Seng Chua, and Qing Li. 2024. A picture is worth a graph: A blueprint debate para- digm for multimodal reasoning. In Proceedings of the 32nd ACM International Conference on Multimedia. 419â428. [179] Andrew Y Zhou, Sharvaree Vadgama, Sumanth Varambally, Peter Eckmann, Michael K Gilson, and Rose Yu. 2026. ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery. arXiv preprint arXiv:2605.12784 (2026). [180]Zhenpeng Zhou et al.2023. QADD: De novo drug design by iterative multi- objective deep reinforcement learning. Bioinformatics 39, 4 (2023), btad157. doi:10.1093/bioinformatics/btad157 [181]Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N. Zare, and Patrick Riley. 2019. Optimization of molecules via deep reinforcement learning. Scientific Reports 9, 1 (2019), 10752. doi:10.1038/s41598-019-47148-x [182]Yunheng Zou, Austin H Cheng, Abdulrahman Aldossary, Jiaru Bai, Shi Xuan Leong, Jorge Arturo Campos-Gonzalez-Angulo, Changhyeok Choi, Cher Tian Ser, Gary Tom, Andrew Wang, et al.2025. El Agente: An autonomous agent for quantum chemistry. Matter 8, 7 (2025).