Paper deep dive
BioProAgent: Neuro-Symbolic Grounding for Constrained Scientific Planning
Yuyang Liu, Jingya Wang, Liuzhenghao Lv, Yonghong Tian
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/20/2026, 3:56:44 AM
Summary
The paper introduces BioProAgent, a neuro-symbolic framework designed to bridge the gap between Large Language Model (LLM) reasoning and physical execution in wet-lab environments. It employs a deterministic Finite State Machine (FSM) to enforce a Design-Verify-Rectify workflow, ensuring hardware compliance and preventing irreversible errors. The system utilizes Semantic Symbol Grounding to reduce context window usage and hierarchical verification for scientific and physical safety. BioProAgent achieves 95.6% physical compliance on the BioProBench benchmark, significantly outperforming baselines like ReAct.
Entities (7)
Relation Signals (5)
BioProAgent → uses → Finite State Machine
confidence 95% · BioProAgent, a neuro-symbolic framework that anchors probabilistic planning in a deterministic Finite State Machine (FSM).
BioProAgent → employs → Semantic Symbol Grounding
confidence 92% · address the context bottleneck inherent in complex device schemas by Semantic Symbol Grounding
BioProAgent → enforces → Design-Verify-Rectify
confidence 90% · enforces a rigorous Design-Verify-Rectify workflow
BioProAgent → evaluatedon → BioProBench
confidence 90% · In the extended BioProBench benchmark, BioProAgent achieves 95.6% physical compliance
BioProAgent → outperforms → ReAct
confidence 90% · BioProAgent achieves 95.6% physical compliance (compared to 21.0% for ReAct)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) have demonstrated significant reasoning capabilities in scientific discovery but struggle to bridge the gap to physical execution in wet-labs. In these irreversible environments, probabilistic hallucinations are not merely incorrect; they can cause equipment damage or experimental failure. We propose BioProAgent, a neuro-symbolic framework that anchors probabilistic planning in a deterministic Finite State Machine (FSM). We introduce a State-Augmented Planning mechanism that enforces a rigorous Design-Verify-Rectify workflow, ensuring hardware compliance before execution. Furthermore, we address the context bottleneck inherent in complex device schemas by Semantic Symbol Grounding, reducing token consumption by ~6* through symbolic abstraction. In the extended BioProBench benchmark, BioProAgent achieves 95.6% physical compliance (compared to 21.0% for ReAct), demonstrating that neuro-symbolic constraints are essential for reliable autonomy in irreversible physical environments. Code: this https URL | Website: this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2603.00876v2
- Canonical: https://arxiv.org/abs/2603.00876v2
Trouble viewing inline? Open PDF directly →
Full Text
85,938 characters extracted from source content.
Expand or collapse full text
BioProAgent: Neuro-Symbolic Grounding for Constrained Scientific Planning Yuyang Liu1,2,† , Jingya Wang2, Liuzhenghao Lv3, Yonghong Tian1,2,3,† 1School of AI for Science, Peking University 2School of Electronic and Computer Engineering, Peking University 3School of Computer Science, Peking University liuyuyang13, yhtian@pku.edu.cn, lvliuzh@stu.pku.edu.cn Abstract Large language models (LLMs) have demonstrated significant reasoning capabilities in scientific discovery but struggle to bridge the gap to physical execution in wet-labs. In these irreversible environments, probabilistic hallucinations are not merely incorrect; they can cause equipment damage or experimental failure. We propose BioProAgent, a neuro-symbolic framework that anchors probabilistic planning in a deterministic Finite State Machine (FSM). We introduce a State-Augmented Planning mechanism that enforces a rigorous Design-Verify-Rectify workflow, ensuring hardware compliance before execution. Furthermore, we address the context bottleneck inherent in complex device schemas by Semantic Symbol Grounding, reducing token consumption by ∼ 6× through symbolic abstraction. In the extended BioProBench benchmark, BioProAgent achieves 95.6% physical compliance (compared to 21.0% for ReAct), demonstrating that neuro-symbolic constraints are essential for reliable autonomy in irreversible physical environments. Code: https://github.com/YuyangSunshine/bioproagent | Website: BioPro-Project. BioProAgent: Neuro-Symbolic Grounding for Constrained Scientific Planning Yuyang Liu1,2,† , Jingya Wang2, Liuzhenghao Lv3, Yonghong Tian1,2,3,† 1School of AI for Science, Peking University 2School of Electronic and Computer Engineering, Peking University 3School of Computer Science, Peking University liuyuyang13, yhtian@pku.edu.cn, lvliuzh@stu.pku.edu.cn Figure 1: Scientific Reasoning vs. Automation Executability. Vanilla LLMs occupy the Theoretical Zone (high logic, weak code), while Neural Agents (like ReAct) often generate better code but fail in scientific reasoning. BioProAgent achieves Trustworthy Autonomy with superior performance in both dimensions. 1 Introduction Large Language Models (LLMs) are increasingly shifting from static knowledge retrieval to active world modeling in scientific discovery. In the Self-Driving Laboratories (SDL) (Stein and Gregoire, 2019; Abolhasani and Kumacheva, 2023), LLMs serve as central cognitive engines. LLMs, operating as Agentic AI Sapkota et al. (2025); Gridach et al. (2025), can design complex chemical synthesis schemes Boiko et al. (2023), discover novel materials Ghafarollahi and Buehler (2025), and act as laboratory assistants Darvish et al. (2025). Recently, multi-agent systems (Jin et al., 2025; Zhou et al., 2025) have demonstrated impressive capabilities in biological reasoning and computational workflow orchestration. However, a critical Execution Gap Wang and et al. (2024) emerges when transitioning from computer simulation predictions to in vitro physical experiments. Despite the remarkable physical autonomy achieved by robotic platforms, from mobile robotic chemists Burger et al. (2020) to autonomous materials synthesis labs Szymanski et al. (2023), they still operate under rigid, task-specific workflows. Unlike reversible code and virtual environments in software engineering, due to irreversible physical laws, any deviation in parameters can lead to catastrophic equipment failures, sample loss, or unreproducible experimental in biological experimental environments. Despite the reasoning capabilities of existing scientific agents, they still face key challenges in high-risk automated applications. First, standard agents Yao et al. (2022); Shinn et al. (2023) suffer from Cognitive Drift, where contextual overload leads to process confusion, resulting in skipped critical steps or confused temporal dependencies. While current memory systems Chhikara et al. (2025); Packer et al. (2023) excel at semantic retrieval, they struggle with lossless tracking of distinct physical entities. A slight “lost-in-the-middle” Liu et al. (2024) confusion between similar reagent IDs can cause experimental failure. More critically, there is a distinct lack of Pre-Execution Interlocks. While previous neuro-symbolic methods like PAL Gao et al. (2023), Program-of-Thoughts Chen et al. (2022), and SayCan Ahn et al. (2022) have combined LLMs with planners to improve logical correctness, they primarily focus on reasoning optimality. They cannot handle the strict and irreversible safety constraints in wet labs because they lack deterministic control mechanisms based on physical rules that stop and self-correct before execution. To address these challenges, we propose a neuro-symbolic BioProAgent for achieving trustworthy autonomy (Fig. 1), which grounds probabilistic reasoning in a deterministic backbone network with a training-free Finite State Machine (FSM). Unlike static agents restricted to pre-defined workflows, BioProAgent flexibly adapts to diverse tasks by using FSM as a safety boundary to enforce a rigorous “Design-Verify-Rectify” workflow. This controller ensures that all hardware instructions must undergo hierarchical verification for both scientific logic and physical safety before being issued. Furthermore, to address context challenges, we introduce Semantic Symbol Grounding. By decoupling the high-dimensional payloads into symbolic pointers, we reduce token consumption by ∼ 6× while ensuring 100% resource consistency. Extensive evaluation on extended BioProBench Liu et al. (2025b) shows that BioProAgent achieves an 90.0% success rate in error recovery, compared to a complete failure (0%) for the standard baseline. Our contributions are summarized as follows: • We propose BioProAgent, a training-free neuro-symbolic framework that bridges the gap between probabilistic reasoning and deterministic physical execution. • We introduce a State-Augmented Planning mechanism driven by a deterministic FSM and a Semantic Symbol Grounding technique that effectively addresses context problem. • Extensive experiments demonstrate BioProAgent achieves state-of-the-art performance in scientific validity and physical consistency, and significantly outperforms baselines. 2 Related Works 2.1 LLM Agents for Scientific Discovery Autonomous research has evolved from basic tool usage (Bran et al., 2023; Boiko et al., 2023) to complex multi-agent collaboration (Zhang et al., 2025; Jin et al., 2025; Huang et al., 2025). Recent frameworks like DeepScientist Weng et al. (2025) and Organa Darvish et al. (2025) demonstrate long-horizon autonomy, and BioMARS Qiu et al. (2025) and AutoLabs Panapitiya et al. (2025) introduce domain-specific visual monitoring and self-correction. AI-native operating systems Sim et al. (2024); Fei et al. (2024); Gao et al. (2025) are also continuously strengthening the execution infrastructure. However, despite these advances, a critical gap remains: the lack of cognitive safety interlocking mechanisms, making general-purpose agents vulnerable to irreversible physical damage. 2.2 Neuro-Symbolic Reasoning & Action Coupling LLMs with symbolic planners or code interpreters, such as PAL Gao et al. (2023), Program-of-Thoughts Chen et al. (2022), and LLM+P Liu et al. (2023), significantly boosts logical and planning accuracy. In embodied settings, SayCan Ahn et al. (2022) and Voyager Wang et al. (2023) ground actions in physical affordances, while iterative frameworks like ReAct Yao et al. (2022) and Reflexion Shinn et al. (2023) improve reasoning via verbal feedback. Recent work Hyun et al. (2025) utilizes physics-aware rejection sampling to align LLM reasoning traces for materials discovery. While these methods are effective in completing tasks in reversible or simulated environments, they lack mechanisms for enforcing irreversible safety constraints. 2.3 Long-Horizon Context Management While Longformer Beltagy et al. (2020) and RAG Lewis et al. (2020) extend context windows, they often fragment the coherent narrative essential for protocol execution. Agent-specific memory systems like MemGPT Packer et al. (2023) and Mem0 Chhikara et al. (2025) maintain multi-session coherence through hierarchical compression. Recent approaches such as ACON Kang et al. (2025) and LLM-State Chen et al. (2023) address long-horizon tasks via learned state compression. However, these purely neural methods remain probabilistic and inherently lossy, making them vulnerable to “lost-in-the-middle” phenomenon Liu et al. (2024). This ambiguity is unacceptable for tracking precise scientific entities like reagent IDs. Figure 2: Overview of BioProAgent. (1) Cognitive Memory utilizes Symbolic Grounding Φ to manage context efficiently; (2) Neural Planner πθ _θ is grounded in a Design-Verify-Rectify FSM Δ(σ) (σ); (3) Hierarchical Verification (s,pK_s,K_p) acts as a safety interlock, enforcing physical compliance by deterministically triggering rectification. 3 Methodology We formulate automated biological protocol generation as a constrained scientific planning problem. Let P denote possible protocol sequences derived from the valid action space A defined by the Hardware Registry Ω . Given an intent I and environmental context ℰE, our goal is to identify the optimal executable protocol P∗∈P^* that is scientifically valid and physically executable, as: P∗=argmaxP∈ℙθ(P|I,ℰ)⋅[s(P)]⋅[p(P)],P^*= P argmax\;P_θ(P|I,E)·I[K_s(P)]·I[K_p(P)], (1) where [⋅]I[·] is indicator function. sK_s and pK_p the Scientific and Physical verification functions, respectively. (Summary of notations in Appendix A.) As illustrated in Fig. 2, our proposed BioProAgent consists of three main modules: a Hybrid Cognitive Memory, a Neuro-Symbolic Core driven by a deterministic FSM, and a Hierarchical Verification framework. 3.1 Hybrid Cognitive Memory Architecture Existing agents often suffer from cognitive drift in long-horizon tasks, where early context regarding reagent IDs is lost. To address this, we design a hierarchical memory system ℳ=⟨ℳwork,ℳepisodic,ℳlong⟩M= _work,M_episodic,M_long , which decouples reasoning logic from high-dimensional data payloads. Structured Working Memory ℳworkM_work. Direct input of raw JSON schemas (often >10k tokens) induces context saturation. We propose Semantic Symbol Grounding, which projects complex data artifacts into symbolic references with lightweight semantic previews. During planning, the working context is compressed via a projection function Φ : Ccontext←Φ(ℳwork)=(ki,preview(vi))∣(ki,vi)∈ℳwork. splitC_context&← (M_work)\\ &=\(k_i,preview(v_i)) (k_i,v_i) _work\. split (2) For execution, a generated symbolic action ata_t is grounded via a resolution function Ψ : atexec←Ψ(at,ℳwork)a_t^exec← (a_t,M_work), which dynamically maps pointers k back to their complete physical parameters v. This reduces processing complexity from (|v|)O(|v|) to (j)O(j) where j≪|v|j |v| is the dimension of the symbolic pointer, ensuring the planner focuses on logic flow rather than data parsing. Episodic and Long-term Memory. To prevent process disorientation, the episodic memory ℳepisodicM_episodic tracks the active FSM-guided trajectory τt=[(s0,a0,σ0),…,(st−1,at−1,σt−1)] _t=[(s_0,a_0, _0),…,(s_t-1,a_t-1, _t-1)], which explicitly records the cognitive states, executed actions, and returned constraint signals. Meanwhile, the long-term memory ℳlongM_long persists cross-session domain knowledge. Rather than statically bloating the prompt, ℳlongM_long is dynamically queried via retrieval tools on demand. Finally, the neural planner synthesizes these active layers to form a unified contextual prompt: Prompt=Retrieved(ℳlong)⊕τt⊕Φ(ℳwork)⊕stPrompt_t=Retrieved(M_long) _t (M_work) s_t, ensuring procedural awareness without context saturation. 3.2 State-Augmented Neuro-Symbolic Plan To address the lack of strict adherence to state-related rules required by safety-critical hardware in probabilistic LLMs, BioProAgent uses a deterministic finite state machine (FSM) to decouple reasoning from control, aiming to enforce a rigorous Design-Verify-Rectify (DVR) workflow. The system is formally defined as FSM=⟨,Σ,,Δ,πθ⟩FSM= , ,A, , _θ . Unlike finite state automata used in traditional robots, which hard-code specific operation sequences (e.g., Move_A_to_B), our states (e.g., DESIGN, VERIFY, RECTIFY) control the cognitive processes of generation and verification, as illustrated in Fig. 2. This abstraction allows BioProAgent to maintain a consistent control architecture across different domains, from molecular cloning to chemical synthesis, without requiring architectural reconstruction. Algorithm 1 BioProAgent FSM-Gated Execution 0: Intent I, Hardware Registry Ω , Knowledge ℳlongM_long 0: Protocol P∗P^* or Failure 1: Initialize: ℳwork←Parse(I)M_work (I), s0←INITs_0 2: Initialize: Trajectory τ←[]τ←[], Signal σ←σ 0, t←0t← 0 3: while st≠SUCCESSs_t and t<Tmaxt<T_max do 4: st←Δ(σ,st−1)s_t← (σ,s_t-1) Refer to Priority Matrix (Tab. 1) 5: if st∈RECTIFY_DRAFT,RECTIFY_CODEs_t∈\RECTIFY\_DRAFT,RECTIFY\_CODE\ then 6: Update prompt via σ 7: end if 8: Generate at∼πθ(a∣st,Φ(ℳwork),τ)a_t _θ(a s_t, (M_work),τ) 9: if ata_t is Draft then 10: Check: s(at)K_s(a_t); Update: σsci∈−1,1 _sci∈\-1,1\ 11: else if ata_t is Code then 12: Check: p(at,Ω)K_p(a_t, ); Update: σphys∈−1,1 _phys∈\-1,1\ 13: if σphys=−1 _phys=-1 then 14: σ←INTERLOCKσ Force Transition via Matrix 15: else 16: Append valid action to P∗P^*: P∗←P∗∪atP^*← P^*∪\a_t\ 17: end if 18: end if 19: Update Memory ℳepisodicM_episodic: τ←τ∪(st,at,σ)τ←τ∪\(s_t,a_t,σ)\ 20: t←t+1t← t+1 21: end while 22: return P∗P^* 3.2.1 Neuro-Symbolic Control Engine To capture both resource availability and validation outcomes, we define the context signal space as Σ=0,13×−1,0,12 =\0,1\^3×\-1,0,1\^2. Let the signal vector at time t be σt=[σknow,σdraft,σcode,σsci,σphys]∈Σ _t=[ _know, _draft, _code, _sci, _phys]∈ . Here, resource signals take boolean values 0,1\0,1\, while validation signals take ternary values −1,0,1\-1,0,1\ (representing Failed, Pending, and Passed, respectively). State transitions Δ:×Σ→ :S× are governed by a deterministic Priority Decision Matrix (Table 1). The physical execution function EphysE_phys for an action ata_t is defined as: Ephys(at)=Execute(at),if at is Code∧(σphys=1)∅,otherwise (Interlock Triggered)E_phys(a_t)= casesExecute(a_t),&if a_t is Code ( _phys=1)\\ ,&otherwise (Interlock Triggered) cases (3) We formalize the system’s reliability as a conditional safety guarantee: P(Unsafe∣Interlock,Ω,Φ)≈0P(Unsafe , , )≈ 0. Interlock denotes the FSM gating mechanism. The guarantee holds under two assumptions: (1) the hardware registry Ω completely covers the physical constraints; and (2) the semantic symbol grounding Φ correctly maps natural language intents to specific device IDs. For unmodeled risks outside Ω , the system conservatively aborts execution and returns a FAILURE signal to ensure hardware safety. The FSM evaluates the ordering rules and selects the target state based on the highest-priority signal match: st+1=Δ(σt,st)=Target(rj)∈s_t+1= ( _t,s_t)=Target(r_j) (4) where rjr_j represents the highest-priority matched rule in Table 1, determined by j=mini∣Match(ri,σt)j= \i (r_i, _t)\. If technical validation fails, the interlock strictly prohibits execution. The complete neuro-symbolic execution, coordinating the interaction between Δ and πθ _θ, is formalized in Algorithm 1. Table 1: FSM Priority Decision Matrix. See Appendix C for the complete definition. Signal Condition (σt _t) Priority Target State (st+1s_t+1) (σphys=−1)∧σcode( _phys=-1) _code 1 RECTIFY_CODE (σsci=−1)∧σdraft( _sci=-1) _draft 2 RECTIFY_DRAFT ¬(σdraft)∧σknow ( _draft) _know 3 DESIGN_DRAFT σdraft∧(σsci=1) _draft ( _sci=1) 4 DESIGN_CODE σdraft∧(σsci=0) _draft ( _sci=0) 5 VERIFY_DRAFT σcode∧(σphys=1) _code ( _phys=1) 6 SUCCESS … … … 3.2.2 Neural Reasoning Policy In each state sts_t, πθ _θ selects an action ata_t from A, corresponding to the set of available tools (Appendix B) categorized into three functional clusters: (1) Design (Generation), (2) Verify (Validation), and (3) Rectify (Correction). To ensure coherence, the FSM dynamically prunes the action space to valid⊂A_valid based on the active phase in sts_t. The planner is conditioned on sts_t, and the system prompt explicitly enforces this state-dependent action masking (Appendix I.1). 3.3 Hierarchical Verification Framework To implement the Verify phase of the DVR loop, we employ a two-layer hierarchical framework to enforce the constraints defined in Eq. (1). Upon execution, the binary outputs of these indicator functions are mapped directly to our FSM’s ternary validation signals (σsci _sci and σphys _phys), where an indicator value of 11 yields a Passed signal (11) and 0 yields a Failed signal (−1-1). Layer 1: Neuro-Scientific Verification (sK_s). The Scientific Reflector assesses the Design draft using prompts derived from expert-in-the-loop iterations. This ensures the CoT reasoning Wei et al. (2022) follows professional wet-lab standards rather than superficial syntax checking. We formalize this evaluation as: [s(Pdraft)]=[Reflector(Pdraft,Task)=“PASS”].I[K_s(P_draft)]=I [Reflector(P_draft,Task)=``PASS′ ]. (5) Full verification criteria, including standard controls and biosafety checks, are detailed in Appendix I.2. Layer 2: Symbolic-Physical Verification (pK_p). To ensure safety, a deterministic Rule Engine (ℛphysR_phys) verifies the generated machine code against the hardware registry Ω . The registry Ω encodes constraints for 22 instruments across Liquid Handling (LH), Thermal Control (TC), and Centrifugation (CF). Crucially, any parameter violation drives the overall product to zero (i.e., [p]=0I[K_p]=0, triggering σphys=−1 _phys=-1), which deterministically activates the Rectify interlock: [p(Pcode)]=∏op∈Pcode∏c∈Ω(deviceop)[Check(op,c)],I[K_p(P_code)]= _op∈ P_code _c∈ (device_op)I[Check(op,c)], (6) where opop is an individual atomic operation within the generated sequence PcodeP_code, deviceopdevice_op denotes the specific hardware instrument targeted by opop, and c represents an individual physical constraint (e.g., maximum speed, temperature limits) defined in the registry Ω for that device. The Boolean function Check(op,c)Check(op,c) returns 11 if the operation complies with the constraint, and 0 otherwise. 4 Experiments 4.1 Experimental Setup Benchmark and Hardware Grounding. We evaluate our framework using an extended version of BioProBench Liu et al. (2025b), including four specialized subsets: Subset A (Protocol Drafting) assesses scientific validity; Subset B (Code Generation) evaluates hardware schema compliance and parameter accuracy; Subset C (Long-Horizon) targets global orchestration, featuring 9 complex protocols with 547 atomic steps (max 238 per task); and Subset D (Error Correction) measures robustness against injected faults. To ensure reproducibility, the complete source code, the digitized hardware registry (Ω ) covering 22 core synthetic biology instruments, and the dataset are provided in the supplementary material. Crucially, although evaluated in simulation, all generated instructions strictly adhere to the API specifications of standard automated devices to bridge the sim-to-real gap. Baselines. We compare BioProAgent against three categories of baselines: (1) Vanilla LLMs: GPT-4o (Islam and Moushi, 2025), Gemini-3-Flash Google (2025), DeepSeek-V3.2 Liu et al. (2025a), and Kimi-k2-instruct Team et al. (2025) via direct prompting; (2) Standard Agents: ReAct Yao et al. (2022), Reflexion Shinn et al. (2023), and AutoGPT Yang et al. (2023); (3) Domain-Specific Agent: Biomni Huang et al. (2025) (evaluated on native protocol generation). Critical Implementation Detail: To ensure fair comparison, all standard agent baselines were equipped with the exact same toolset as BioProAgent. Table 2: The BioProAgent Evaluation Framework. Metrics are categorized by dimension, improvement direction (↑ /↓ ), and evaluation source. Dimension Evaluator Key Metrics Scientific Semantic Match SsemS_sem (↑ ), ROUGE-L (↑ ) LLM-as-a-Judge CsC_s (↑ ) Physical Rule Engine ScodeS_code (↑ ), CpC_p (↑ ) Ground Truth AccseqAcc_seq (↑ ), AccparamAcc_param (↑ ) Efficiency System Logs Succ.Succ. (↑ ), Tokens (↓ ), Loop Rate (↓ ) Evaluation Framework We evaluate performance across three dimensions: (1) Scientific Validity: Measured by a composite Semantic Score SsemS_sem for keyword coverage and an LLM-as-a-Judge Scientific Score CsC_s for procedural logic. (2) Physical Compliance: Quantified by the Overall Code Score ScodeS_code, which aggregates Physical Compliance CpC_p verified by the symbolic rule engine, along with Sequence AccseqAcc_seq and Parameter Accuracy AccparamAcc_param. Notably, ScodeS_code acts as a hard gatekeeper: schema format failures yield a zero score. (3) System Efficiency: Tracks Success Rate Succ.Succ., token consumption Tokens, and Loop Rate. Table 2 summarizes these metrics; mathematical derivations are detailed in Appendix E. Table 3: Main Results on Protocol Drafting (Subset A) and Code Generation (Subset B). † : indicates statistical significance compared to all baselines (Biomni, ReAct, AutoGPT, Reflexion) with p<0.001p<0.001 (paired t-test). Subset A: Scientific Reasoning Subset B: Hardware Execution Method Backbone ROUGE-L ↑ Ssem↑S_sem Cs†↑C_s [rgb]0,1,0 [named]pgfstrokecolorrgb0,1,0 Time (s) ↓ Scode↑S_code Cp↑C_p Accparam↑Acc_param Direct GPT-4o 0.107 0.202 0.189 13.8 0.590 0.995 0.295 Direct Gemini-3-Flash 0.130 0.247 0.322 12.1 0.576 0.996 0.287 Direct DeepSeek-V3 0.123 0.260 0.285 52.1 0.495 0.995 0.205 Direct Qwen3-30B-a3b 0.124 0.249 0.201 20.0 0.509 0.988 0.222 Direct Kimi-k2-Instruct 0.097 0.271 0.286 83.1 0.499 0.971 0.215 Biomni (Specialized) 0.081 0.252 0.342 87.1 N/A N/A N/A ReAct Gemini-3-Flash 0.116 0.268 0.455 44.5 0.038 0.210 0.103 Reflexion Gemini-3-Flash 0.118 0.282 0.439 148.4 0.278 0.534 0.403 AutoGPT Gemini-3-Flash 0.116 0.258 0.429 119.6 0.540 0.911 0.468 BioProAgent Gemini-3-Flash 0.147 0.344 0.591 71.8 0.653 0.956 0.610 Table 4: Main Results on Long-Horizon Stability (Subset C) and Error Correction (Subset D). Bold indicates the best performance within agent-based frameworks. Subset C: Long-Horizon Stability Subset D: Error Correction Method Backbone Succ. ↑ AccparamAcc_param ↑ CpC_p ↑ Correction Succ. ↑ CpC_p ↑ Loop Rate ↓ ReAct Gemini-3-Flash 88.9% 0.114 0.217 0.0% 0.000 40.0% Reflexion Gemini-3-Flash 33.3% 0.000 0.000 0.0% 0.000 0.0% AutoGPT Gemini-3-Flash 66.7% 0.409 0.644 0.0% 0.000 0.0% BioProAgent Gemini-3-Flash 100.0% 0.718 0.950 90.0% 0.925 0.0% 4.2 Main Results 4.2.1 Scientific Reasoning (Subset A). As shown in Table 3, BioProAgent achieves a Scientific Validity Score (CsC_s) of 0.591, a 30% improvement over the strongest agent baseline (ReAct) and almost double that of vanilla models (Gemini-3-Flash). Notably, even the domain-specialized Biomni only reaches a CsC_s of 0.342, highlighting that domain knowledge alone is insufficient without structured reasoning constraints. Paired t-tests confirm that these improvements are statistically significant (p<0.001p<0.001). BioProAgent’s advantage stems from its Scientific Reflector within the FSM, which requires a formal review of scientific logic before proceeding to the coding phase, whereas standard agents often neglect crucial controls in pursuit of rapid generation. Cross-model validation (Pearson r≥0.84r≥ 0.84) confirms the robustness of our LLM-as-a-Judge metric (see Appendix E.4.2). 4.2.2 Physical Compliance (Subset B). Table 3 reveals significant differences in code generation performance between the baselines. Vanilla LLMs maintain Cp=0.995C_p=0.995 primarily by defaulting to conservative, low-complexity code, which naturally results in poor Parameter Accuracy (0.295). Standard Agents (e.g., ReAct), conversely, attempt complex operations but suffer from catastrophic safety failures (Cp=0.210C_p=0.210). Although ReAct was equipped with the exact same Rule Engine tool, it often fails to invoke these tools proactively due to context saturation or a lack of stopping conditions. BioProAgent effectively eliminates this trade-off, achieving the highest ScodeS_code and AccparamAcc_param while maintaining high-quality security (Cp=0.956C_p=0.956). This confirms FSM’s deterministic gating is essential for enforcing safety checks that probabilistic agents tend to bypass. (a) (b) (c) Figure 3: Efficiency and Reliability Analysis. (a) Multidimensional Capabilities: BioProAgent (red) achieves the most expansive capability envelope across five critical dimensions. Note that identical metrics on different axes represent performance in distinct evaluation subsets. (b) Cost-Efficiency: On long-horizon tasks (Subset C), our system reduces token consumption by 82% compared to AutoGPT while maintaining a 100% success rate. (c) Time-Efficiency: BioProAgent demonstrates superior precision with concise execution time. 4.2.3 Long-Horizon Stability (Subset C). As shown in Table 4, standard agents exhibit significant performance gaps when the workflow expands to an average of 60 steps. While AutoGPT maintains a 66.7% success rate, its lack of a structural stopping condition leads to inefficient thought loops and excessive token consumption. ReAct achieves an 88.9% task completion rate, but its low physical consistency and Accparam=0.114Acc_param=0.114 reveal ID drift. Notably, Reflexion fails completely, resulting in Accparam=0.000Acc_param=0.000. This is attributed to “Schema Corruption”: its verbal feedback mechanism frequently injects conversational artifacts (e.g., apologies) into the JSON payload, causing the strict rules engine to fail to parse the output. In contrast, by replacing the high-dimensional data payloads with symbolic pointers, our Semantic Symbol Grounding ensures 100% resource consistency and lossless state tracking. 4.2.4 Error Correction (Subset D). Subset D evaluates the system’s ability to diagnose and correct injected physical violations. We define a successful correction as resolving all HALT-level safety threats (Cp>0.8C_p>0.8). As shown in Table 4, all standard baseline agents exhibit a 0% correction rate, with ReAct entering infinite retry loops in 40% of scenarios. To determine if this is merely a tool-use policy failure, we introduced an Oracle-ReAct variant with a forced interlock, prohibiting termination until a “PASS” verification signal was received. Oracle-ReAct still achieved a 0% success rate, falling into a 100% infinite loop. Without Semantic Symbol Grounding, attempting to feed massive, broken JSON payloads back into the reasoning window caused severe “Context Paralysis.” In contrast, BioProAgent successfully restores physical safety (CpC_p) to 0.925 and achieves an outstanding 90.0% correction success rate. The neuro-symbolic FSM acts as a deterministic router: when a violation is detected by the rules engine, it overwrites the LLM’s trajectory and forces a transition to the RECTIFY_CODE state, aligning outputs with hardware limitations. Appendix F.2 provides a detailed analysis of logical errors, resource errors, and physical errors. Furthermore, detailed analysis evaluating whether scaling up to massive “Thinking” models can inherently solve this execution gap is provided in Appendix G, further proving the necessity of our framework. 4.3 Holistic Capability Assessment Fig. 3(a) visualizes the capability envelope of each agent. Crucially, this chart aggregates performance across diverse benchmarks: while CpC_p and ScodeS_code appear on multiple axes, they correspond to distinct stress tests, standard generation (Subset B), long-horizon stability (Subset C), and error correction (Subset D). Existing methods exhibit significant capability biases: Direct Prompting (Grey) defaults to conservative safety (CpC_p) but fails in complex coordination, while ReAct (Green) improves scientific reasoning (CsC_s) at the cost of severe physical hallucinations. BioProAgent effectively eliminates this trade-off, achieving a balanced pentagonal coverage. The performance gap is most significant in Long-Horizon and Self-Correction dimensions. Baselines degrade to near-zero performance due to context collapse or open-loop error propagation, BioProAgent maintains robustness, demonstrating that neuro-symbolic FSM is essential for sustaining trustworthy autonomy in high-stakes environments. 4.4 Efficiency and Cost Analysis As shown in Fig. 3(b), BioProAgent consumes 82% fewer tokens than the AutoGPT on Subset C. This efficiency stems from our Semantic Symbol Grounding mechanism, which maintains a near-constant context size by substituting dense data payloads with symbolic pointers, thereby mitigating the "context explosion" inherent in processing dense JSON payloads . Furthermore, Fig. 3(c) highlights the critical trade-off between execution time and precision. While standard baselines are faster, they lack the requisite precision for complex hardware operations. While the AutoGPT attains moderate precision, its execution time is prohibitively long (avg. 695s). BioProAgent achieves an optimal balance, delivering high-precision results with competitive execution speeds and the widest safety margin (CpC_p), validating practical viability for the real-world laboratory deployment. (a) Physical Violation (V>VmaxV>V_max)DESIGNCODEOutput ata_t:25,000gRECTIFYCODEInterlock(σphys− _phys^-)DESIGNCODERevised at′a_t :15,000gSUCCESSPass pK_p(σphys+ _phys^+)REJECTFeedbackACCEPT(b) Symbol Grounding Error (Undefined ID)DESIGNCODEOutput ata_t:"new_plate"RECTIFYCODEGrounding Fail(Lookup Error)DESIGNCODEResolved:"plate_1"SUCCESSValid ∈Ω∈ Figure 4: FSM-Driven Self-Correction Trajectories. (a) Physical Correction. (b) Symbol Grounding Repair. 4.5 Qualitative Analysis Mechanism of Correction. Standard LLM agents typically operate in an open-loop manner: when a hallucinated or dangerous parameter is generated , execution is immediate, inevitably leading to failure. In contrast, Fig. 4 illustrates the FSM’s correction mechanism in action, effectively intercepting speed limit violations. In the physical violation case (Fig. 4a), the deterministic Rule Engine (ℛphysR_phys) preemptively intercepts this violation (V>VmaxV>V_max), triggering a forced state transition to RECTIFY_CODE. This forces the Neural Planner to recognize the specific error signal (σphys=−1 _phys=-1) and regenerate the code within safe limits (15,000g). Similarly, for a logical hallucination case (Fig. 4b), the system detects semantic drift (an undefined resource ID new_plate). The FSM prevents this error propagation, guiding the agent to remap the action to the grounded symbol plate_1. Table 5: Representative Error Corrections in Subset D. The FSM interceptor enforces the Rectify phase, guiding the agent from violations to valid execution. Constraint Design Violation (sdesigns_design) FSM Rectification (srectifys_rectify) Physical (pK_p) centrifuge(speed=‘25000g’) centrifuge(speed=‘15000g’) Trace × Exceeds rotor limit (15k) ✓ Clamped to safety max Symbol (Φ ) transfer(source=‘buffer_x’) transfer(source=‘trough_1’) Trace × Grounding Fail (No ID) ✓ Remapped to registry Causal seal_plate(); transfer(...) transfer(...); seal_plate() Trace × Blocked op (Sealed) ✓ Reordered dependency Granularity of Rectification. Table 5 further demonstrates that these corrections are not merely stochastic retries but involve semantic-level reasoning. We observe three distinct categories of rectification within the DVR loop: (1) Safety Alignment: Adjusting physics parameters to satisfy the hardware constraints defined in the registry (Ω ); (2) Resource Grounding: Resolving hallucinated variables to physically validated slots via the projection function Φ ; and (3) Causal Rectification: Correcting logical dependencies, such as the critical reordering of seal_plate() operations. This confirms that BioProAgent’s neuro-symbolic core effectively operationalizes a “Design-Verify-Rectify” workflow, ensuring trustworthy execution in high-stakes environments. Step01020304050BioProAgentPlanCodeSuccess(Step 8)ReActRetrieve×ContextOverflowAutoGPTRedundant Info-Seeking Loop×Timeout(Step 45)PlanCodeRetrievalLoop Figure 5: Trace Analysis (Subset C). Execution divergence on a long-horizon task. TOP: BioProAgent efficiently transitions to coding via FSM. MIDDLE: ReAct fails early due to context overflow. BOTTOM: AutoGPT wastes resources in a retrieval loop. Log-Level Behavioral Analysis. To diagnose why baseline agents fail in long-horizon tasks despite having access to the same tools, we visualize the execution trajectories of a representative metabolomics workflow in Fig. 5. ReAct terminates prematurely (Step 7) as retrieved API schemas saturate its context window, blocking valid action generation. Lacking a deterministic stop signal, AutoGPT enters an infinite loop of redundant retrieve_knowledge calls (Steps 20–45), consuming excessive tokens without advancing state. In contrast, BioProAgent’s FSM enforces a rigid Plan → Code transition, acting as a cognitive clock that ensures efficient completion in just 8 steps. 4.6 Ablation Study Table 6 quantitatively validates the architectural necessity of each component. Impact of FSM: Removing the FSM controller leads to a structural failure. The Scientific Validity (CsC_s) degrades to 0.262, while performance on long-horizon tasks (Subset C) collapses to 0.000. Crucially, its physical safety on error-correction tasks (Subset D) drops to 0.782. It is important to note that 0.782 is not a partial success, but rather the baseline physical compliance of the corrupted inputs themselves. Because the agent lacks the deterministic state routing (i.e., RECTIFY_CODE) provided by the FSM, it fails to execute any meaningful self-correction, blindly translating the injected hazards into executable code. Impact of Verification: Ablating the verifications (p,sK_p,K_s) causes CpC_p to drop from 0.956 to 0.902. While less dramatic than the FSM ablation, this deficit demonstrates that neural inference alone, even with advanced prompting, cannot guarantee the zero-defect safety required for irreversible wet-lab automation. Impact of Knowledge & Disambiguation: Omitting the disambiguation tool (clarify_experiment_scope) or external knowledge retrieval (retrieve_knowledge) significantly impairs reasoning capabilities. Notably, the sharp decline in CsC_s (0.591→0.4030.591→ 0.403) without clarification underscores the critical role of human-in-the-loop disambiguation for handling under-specified protocols. Statistical significance is further confirmed via cross-judge verification (Appendix E.4.1) and paired t-tests (Appendix E.4.2). Table 6: Ablation Study of BioProAgent Components. Subset A Subset B Subset C Subset D Variant CsC_s CpC_p ScodeS_code CpC_p Full Model 0.591 0.956 0.668 0.925 w/o FSM 0.262 0.049 0.000 0.782 w/o Verification 0.563 0.902 0.648 0.782 w/o Knowledge 0.561 0.889 0.657 N/A w/o Clarify 0.403 N/A 0.632 N/A 5 Conclusion We present BioProAgent, a neuro-symbolic framework addressing the critical execution gap in autonomous scientific discovery. By grounding probabilistic LLM reasoning within a deterministic Finite State Machine, BioProAgent enforces a “Design-Verify-Rectify” workflow for irreversible wet-lab environments. Evaluations on extended BioProBench demonstrate that BioProAgent achieves SOTA performance and demonstrates robust autonomous self-correction where traditional agents fail. Our work emphasizes that for high-risk physical applications, neural intelligence must be constrained by symbolic safety interlocks to ensure trustworthy and reliable autonomy. Limitations & Future Work Despite its performance, BioProAgent has some limitations. First, the framework’s security relies on a predefined hardware registry (Ω ). While effective for known instruments, this requires manual registration for custom hardware. Future work will explore using a multimodal Large Language Model (MLLM) to automate this process by directly reading the device manual. Additionally, our current evaluation was conducted in high-fidelity simulations using standard APIs. While this validates the correctness of the logic and parameters, real-world physical randomness remains unmodeled. Integrating a real-time visual feedback loop is crucial for managing these physical variables. Acknowledgement Supported by the China Postdoctoral Science Foundation under Grant Numbers BX20240013 and 2024M760113, the Shenzhen Science and Technology Innovation Commission under Grant KQTD 20240729102051063, and the National Natural Science Foundation of China under Grant Numbers U25B6003 and 62425101. References Abolhasani and Kumacheva (2023) Milad Abolhasani and Eugenia Kumacheva. 2023. The rise of self-driving labs in chemical and materials sciences. Nature Synthesis, 2(6):483–492. Ahn et al. (2022) Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, and 1 others. 2022. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691. Beltagy et al. (2020) Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. Boiko et al. (2023) Daniil A Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. 2023. Autonomous chemical research with large language models. Nature, 624(7992):570–578. Bran et al. (2023) Andres M Bran, Sam Cox, Andrew D White, and Philippe Schwaller. 2023. Chemcrow: Augmenting large-language models with chemistry tools. arXiv preprint arXiv:2304.05376. Burger et al. (2020) Benjamin Burger, Phillip M Maffettone, Vladimir V Gusev, Catherine M Aitchison, Yang Bai, Xiaoyan Wang, Xiaobo Li, Ben M Alston, Buyi Li, Rob Clowes, and 1 others. 2020. A mobile robotic chemist. Nature, 583(7815):237–241. Chen et al. (2023) Siwei Chen, Anxing Xiao, and David Hsu. 2023. Llm-state: Open world state representation for long-horizon task planning with large language model. arXiv preprint arXiv:2311.17406. Chen et al. (2022) Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W Cohen. 2022. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. arXiv preprint arXiv:2211.12588. Chhikara et al. (2025) Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413. Darvish et al. (2025) Kourosh Darvish, Marta Skreta, Yuchi Zhao, Naruki Yoshikawa, Sagnik Som, Miroslav Bogdanovic, Yang Cao, Han Hao, Haoping Xu, Alán Aspuru-Guzik, Animesh Garg, and Florian Shkurti. 2025. Organa: A robotic assistant for automated chemistry experimentation and characterization. Preprint, arXiv:2401.06949. Fei et al. (2024) Yuxing Fei, Bernardus Rendy, Rishi Kumar, Olympia Dartsi, Hrushikesh P Sahasrabuddhe, Matthew J McDermott, Zheren Wang, Nathan J Szymanski, Lauren N Walters, David Milsted, and 1 others. 2024. Alabos: a python-based reconfigurable workflow management framework for autonomous laboratories. Digital Discovery, 3(11):2275–2288. Gao et al. (2025) Jing Gao, Junhan Chang, Haohui Que, Yanfei Xiong, Shixiang Zhang, Xianwei Qi, Zhen Liu, Jun-Jie Wang, Qianjun Ding, Xinyu Li, and 1 others. 2025. Unilabos: An ai-native operating system for autonomous laboratories. arXiv preprint arXiv:2512.21766. Gao et al. (2023) Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: Program-aided language models. In International Conference on Machine Learning, pages 10764–10799. PMLR. Ghafarollahi and Buehler (2025) Alireza Ghafarollahi and Markus J Buehler. 2025. Sciagents: automating scientific discovery through bioinspired multi-agent intelligent graph reasoning. Advanced Materials, 37(22):2413523. Google (2025) Google. 2025. Gemini 3 flash model card. Online. Google and Google DeepMind (2025) Google and Google DeepMind. 2025. Gemini 3 pro. Online. Gridach et al. (2025) Mourad Gridach, Jay Nanavati, Khaldoun Zine El Abidine, Lenon Mendes, and Christina Mack. 2025. Agentic ai for scientific discovery: A survey of progress, challenges, and future directions. arXiv preprint arXiv:2503.08979. Grootendorst (2020) Maarten Grootendorst. 2020. Keybert: Minimal keyword extraction with bert. Huang et al. (2025) Kexin Huang, Serena Zhang, Hanchen Wang, Yuanhao Qu, and 1 others. 2025. Biomni: A general-purpose biomedical ai agent. bioRxiv. Preprint. Hyun et al. (2025) Lee Hyun, Sohee Yoon, Jinwoo Park, Sue In Chae, Seongeon Park, Jooyeon Ahn, Yebin Jung, Youjung Chung, Hogeun Chang, Sujin Park, and 1 others. 2025. Aligning reasoning llms for materials discovery with physics-aware rejection sampling. arXiv preprint arXiv:2509.00768. Islam and Moushi (2025) Raisa Islam and Owana Marzia Moushi. 2025. Gpt-4o: The cutting-edge advancement in multimodal llm. In Intelligent Computing-Proceedings of the Computing Conference, pages 47–60. Springer. Jin et al. (2025) Ruofan Jin, Yucheng Guo, Yuanhao Qu, Ming Yang, and 1 others. 2025. Biolab: End-to-end autonomous life sciences research with multi-agents system integrating biological foundation models. bioRxiv. Preprint. Kang et al. (2025) Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, and Saravan Rajmohan. 2025. Acon: Optimizing context compression for long-horizon llm agents. arXiv preprint arXiv:2510.00615. Lewis et al. (2020) Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459–9474. Liu et al. (2025a) Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025a. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556. Liu et al. (2023) Bo Liu, Yuqian Jiang, Xiaohan Zhang, Qiang Liu, Shuo Zhang, Joydeep Biswas, and Peter Stone. 2023. Llm+p: Empowering large language models with optimal planning proficiency. arXiv preprint arXiv:2304.11477. Liu et al. (2024) Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173. Liu et al. (2025b) Yuyang Liu, Liuzhenghao Lv, Xiancheng Zhang, Li Yuan, and Yonghong Tian. 2025b. Bioprobench: Comprehensive dataset and benchmark in biological protocol understanding and reasoning. arXiv preprint arXiv:2505.07889. Packer et al. (2023) Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. 2023. Memgpt: Towards llms as operating systems. Panapitiya et al. (2025) Gihan Panapitiya, Emily Saldanha, Heather Job, and Olivia Hess. 2025. Autolabs: Cognitive multi-agent systems with self-correction for autonomous chemical experimentation. arXiv preprint arXiv:2509.25651. Qiu et al. (2025) Yibo Qiu, Zan Huang, Zhiyu Wang, Handi Liu, Yiling Qiao, Yifeng Hu, Shu’ang Sun, Hangke Peng, Ronald X Xu, and Mingzhai Sun. 2025. Biomars: A multi-agent robotic system for autonomous biological experiments. arXiv preprint arXiv:2507.01485. Sapkota et al. (2025) Ranjan Sapkota, Konstantinos I Roumeliotis, and Manoj Karkee. 2025. Ai agents vs. agentic ai: A conceptual taxonomy, applications and challenges. arXiv preprint arXiv:2505.10468. Shinn et al. (2023) Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652. Sim et al. (2024) Malcolm Sim, Mohammad Ghazi Vakili, Felix Strieth-Kalthoff, Han Hao, Riley J Hickman, Santiago Miret, Sergio Pablo-García, and Alán Aspuru-Guzik. 2024. Chemos 2.0: An orchestration architecture for chemical self-driving laboratories. Matter, 7(9):2959–2977. Stein and Gregoire (2019) Helge S Stein and John M Gregoire. 2019. Progress and prospects for accelerating materials science with automated and autonomous workflows. Chemical science, 10(42):9640–9649. Szymanski et al. (2023) Nathan J Szymanski, Bernardus Rendy, Yuxing Fei, Rishi E Kumar, Tanjin He, David Milsted, Matthew J McDermott, Max Gallant, Ekin Dogus Cubuk, Amil Merchant, and 1 others. 2023. An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature, 624(7990):86. Team et al. (2025) Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534. Wang et al. (2023) Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291. Wang and et al. (2024) Xiaoxuan Wang and et al. 2024. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. arXiv preprint arXiv:2307.10635. Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837. Weng et al. (2025) Yixuan Weng, Minjun Zhu, Qiujie Xie, Qiyao Sun, Zhen Lin, Sifan Liu, and Yue Zhang. 2025. Deepscientist: Advancing frontier-pushing scientific findings progressively. arXiv preprint arXiv:2509.26603. Yang et al. (2023) Hui Yang, Sifu Yue, and Yunzhong He. 2023. Auto-gpt for online decision making: Benchmarks and additional opinions. arXiv preprint arXiv:2306.02224. Yao et al. (2022) Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations. Zhang et al. (2025) Zhen Zhang, Zhichu Ren, Chia-Wei Hsu, Weibin Chen, Zhang-Wei Hong, Chi-Feng Lee, Aubrey Penn, Hongbin Xu, Daniel J Zheng, Shuhan Miao, and 1 others. 2025. A multimodal robotic platform for multi-element electrocatalyst discovery. Nature, pages 1–3. Zhou et al. (2025) Yuyang Zhou, Jin Su, Jiawei Zhang, Wangyang Hu, Tianli Tao, Guanqi Li, Xibin Zhou, Li Fan, and Fajie Yuan. 2025. Prime: A multi-agent environment for orchestrating dynamic computational workflows in protein engineerings. bioRxiv, pages 2025–09. Appendix Table 7: Summary of Notations and Definitions. Symbol Definition Description Problem Formulation I Natural Language Intent The initial user query describing the experimental objective. ℰE Environmental Context The set of available lab hardware, reagents, and physical conditions. P Protocol Space The universal set of all possible protocol sequences. P∗P^* Optimal Protocol The target executable sequence that maximizes scientific validity and physical compliance. ℙθP_θ Generation Probability The probability distribution of generating the complete protocol sequence P. sK_s Scientific Verification Function checking experimental logic constraints (enforced by the Scientific Reflector). pK_p Physical Verification Function checking hardware constraints (enforced by the Rule Engine ℛphysR_phys). [⋅]I[·] Indicator Function Returns 1 if the logical condition inside holds, 0 otherwise. Cognitive Memory & Grounding ℳM Cognitive Memory The hierarchical memory system ⟨ℳwork,ℳepisodic,ℳlong⟩ _work,M_episodic,M_long . Φ Projection Function Maps high-dimensional artifacts to symbolic references with semantic previews. Ψ Resolution Function Dynamically grounds symbolic pointers back to physical parameters (atexec←Ψ(at)a_t^exec← (a_t)). CcontextC_context Compressed Context The lightweight context representation after applying Φ . τt _t Execution Trajectory The history of FSM-guided states, executed actions, and signals up to time t: [(s0,a0,σ0),…][(s_0,a_0, _0),…]. Neuro-Symbolic Core (FSM) S State Space The set of discrete cognitive states (e.g., Design, Verify, Rectify). sts_t Current State The active state at time t, determining the valid toolset validA_valid. Σ Context Signal Space The mixed boolean/ternary signal space defined as 0,13×−1,0,12\0,1\^3×\-1,0,1\^2. σt _t Context Signal Vector A vector [σknow,σdraft,σcode,σsci,σphys]∈Σ[ _know, _draft, _code, _sci, _phys]∈ driving state transitions. Δ Transition Function The deterministic mapping from current signals and states to the next DVR phase: Δ:Σ×→ : ×S . πθ _θ Neural Policy The LLM-based single-step planner policy distribution over actions. A Action Space The complete set of atomic tools available to the agent. Ω Hardware Registry Structured database of device limits (e.g., max RPM, temperature) defining pK_p. Appendix A Table of Notations Table 7 provides a comprehensive reference for the mathematical notations and symbols employed throughout this paper. To facilitate readability, the symbols are categorized by their functional roles within the BioProAgent framework: Problem Formulation, Cognitive Memory, and the Neuro-Symbolic Core. Table 8: The complete toolset definition. Tools are clustered by the function in the Design-Verify-Rectify (DVR) workflow. Note that verification tools explicitly implement the sK_s and pK_p functions to drive FSM state transitions. Tool Name Parameters Function Description Phase I: Design (Generation & Alignment) generate_scientific_draft query, exp_info, knowledge Generates a natural language experimental protocol draft based on user intent. align_draft_to_automation draft, exp_info Maps the approved draft to specific hardware operations (Intermediate step). generate_machine_code aligned_protocol, suggestion Synthesizes executable JSON machine code. Accepts suggestions from the Rectify phase. Phase I: Verify (Hierarchical Validation) reflect_on_protocol protocol_text, query (Scientific Reflector sK_s) Reviews the draft for logic/controls via CoT. Output updates the ternary signal σsci∈−1,1 _sci∈\-1,1\. validate_machine_code exp_flow_json (Physical Rule Engine ℛphysR_phys) Evaluates pK_p by checking JSON against the registry Ω . Output updates the ternary signal σphys∈−1,1 _phys∈\-1,1\. Phase I: Rectify (Self-Correction) modify_protocol protocol, request (Logic Rectification) Revises the draft based on sK_s feedback (triggered when σsci=−1 _sci=-1). fix_machine_code machine_code, errors (Physical Rectification) Autonomously repairs parameters based on error logs from ℛphysR_phys (when σphys=−1 _phys=-1). Auxiliary & Context clarify_experiment_scope query, doc_content Resolves ambiguities in initial request before plan begins. retrieve_knowledge query, keywords Fetches RAG-based literature/protocols to populate context. ask_user_confirmation question Requests explicit approval before critical execution steps. Table 9: The Priority Decision Matrix defining the FSM transition logic. (σphys=−1)( _phys=-1) denotes a physical validation failure, while (σsci=1)( _sci=1) denotes a scientific validation success. Note how the [Rectify] interlocks (Rows 1-2) take strict precedence over [Design] steps. Priority Signal Condition (σt _t) Target State (st+1s_t+1) Logic Description (DVR Flow) 1 (σphys=−1)∧σcode( _phys=-1) _code RECTIFY_CODE [Rectify] Safety Interlock: If Rule Engine (pK_p) detects a physical violation (Severity: HALT), immediately force a fix. 2 (σsci=−1)∧σdraft( _sci=-1) _draft RECTIFY_DRAFT [Rectify] Scientific Review: If Reflector (sK_s) rejects the protocol logic (e.g., missing controls), require revision. 3 ¬σdraft∧σknow _draft _know DESIGN_DRAFT [Design] Knowledge Ready: If knowledge is retrieved but no draft exists, enter drafting mode. 4 σdraft∧(σsci=1) _draft ( _sci=1) DESIGN_CODE [Design] Code Generation: Synthesize executable machine code directly from the verified and approved scientific draft. 5 σdraft∧(σsci=0) _draft ( _sci=0) VERIFY_DRAFT [Verify] Draft Review: Evaluate the newly generated protocol draft against scientific principles (Pending state). 6 σcode∧(σphys=1) _code ( _phys=1) SUCCESS [Finish] Completion: Code exists and passes all physical validations (pK_p). Task complete. 7 σambiguous _ambiguous CLARIFY_INTENT [Aux] Ambiguity Check: If initial intent is unclear, loop back to ask user for parameters. 8 (Default) sts_t (Self-loop) [Wait]: If no condition matches, maintain current state (e.g., during retrieval or accumulation). Appendix B Detailed Toolset To ensure fairness and reproducibility, all agents in our experiments are equipped with the identical set of operational tools. Table 8 details the specifications, categorized by their functional role in the Design-Verify-Rectify (DVR) workflow. Appendix C FSM Decision Logic The Finite State Machine (FSM) acts as the deterministic backbone of BioProAgent, orchestrating the DVR loop. Table 9 details the Priority Decision Matrix used to compute state transitions st+1=Δ(σt,st)s_t+1= ( _t,s_t). The rules are evaluated in descending order of priority (1 is highest), ensuring that safety interlocks always override generative actions. Table 10: Representative Action Schemas from the Hardware Registry (Ω ). The Physical Rule Engine (pK_p) validates generated JSON code against these parameter constraints and physical interlocks. Device Category Action Primitive Parameters (Type: Constraint) Physical Rules (Examples) Centrifugation Centrifuge ∙ speed_g (int: 500−15000500--15000) ∙ time (duration: H:M:S) ∙ temp_C (int: 4−404--40) ∙ brake (enum: off, slow, fast) 1. Speed must not exceed rotor limit (V≤VmaxV≤ V_max). 2. Temperature requires pre-cooling if T<25∘T<25 C. Thermal Control Incubate ∙ temp_C (int: 4−994--99) ∙ shake_rpm (int: 0−20000--2000) ∙ duration (time) 1. Shaking prohibited if plate is unsealed. 2. Max temp depends on plate material (PS vs. P). Liquid Handling Transfer ∙ source (ID string) ∙ dest (ID string) ∙ volume_uL (float: 0.5−10000.5--1000) ∙ tip_type (enum: p20, p300, p1000) 1. Source volume must be >> aspiration volume + dead volume. 2. Tip type must match volume range. PCR Cycling Thermocycle ∙ lid_temp (int: 95−10595--105) ∙ stages (list of [temp, time, cycles]) 1. Lid temp must be >> reaction temp to prevent condensation. Appendix D Hardware Registry and Action Space Schema To support reproducibility and the reconstruction of the Physical Rule Engine (pK_p), we formally define the device action space. The Hardware Registry (Ω ) contains 22 specific instruments. Every action ata_t generated by the agent must strictly conform to the parameterized schemas defined below. D.1 Device Categories and Functions The registry organizes hardware into four functional categories to enable end-to-end wet-lab automation: • Liquid Handling: Automated pipetting workstations, acoustic liquid handlers, and dispensers. • Thermal Control: PCR thermal cyclers, incubators, and plate sealers. • Separation & Processing: Centrifuges, shakers, and magnetic separators. • Analysis: Multi-mode plate readers, qPCR systems, and electrophoresis imagers. D.2 Action Primitives and Parameter Constraints Table 10 specifies the schemas for representative action primitives. All constraints (e.g., integer ranges, enumerations) are strictly enforced by the Physical Rule Engine (pK_p) during the Verification phase of the DVR loop. D.3 Full Registry Availability Due to space constraints, Table 10 presents a subset of the 22 devices. The complete JSON schema definitions, including configurations for the Automated Nitrogen Evaporator, Microbial Colony Picker, Plate Washer, are provided in the codebase. Appendix E Detailed Evaluation Metrics To rigorous assess the quality of generated protocols and executable machine code, we implement a hybrid evaluation framework that integrates semantic similarity analysis, deterministic rule-based verification, and expert-aligned LLM-as-a-Judge assessments. E.1 Protocol Generation Metrics Let Pgt=p1,…,pkP_gt=\p_1,...,p_k\ denote the ground truth protocol sequence, and Pgen=p1′,…,pm′P_gen=\p _1,...,p _m\ denote the generated sequence. Text and Semantic Metrics. We employ three complementary metrics to measure lexical coverage and structural alignment: 1. Keyword F1 (F1KF1_K): Utilizing KeyBERT Grootendorst (2020), we extract the top-k (k=64k=64) domain-specific entities (e.g., reagents, instruments). Precision (PKP_K) and Recall (RKR_K) are computed based on the entity overlap between PgtP_gt and PgenP_gen: F1K=2⋅PK⋅RKPK+RK.F1_K= 2· P_K· R_KP_K+R_K. (7) 2. Step Recall (SRSR) Liu et al. (2025b): To quantify structural completeness, we measure the proportion of ground truth steps that are semantically recovered in the generation: SR=∑s∈Pgt[maxs′∈PgenSim(s,s′)≥τ]|Pgt|,SR= _s∈ P_gtI [ _s ∈ P_genSim(s,s )≥τ ]|P_gt|, (8) where Sim(⋅)Sim(·) denotes the cosine similarity of sentence embeddings derived from all-mpnet-base-v2, with a semantic threshold τ=0.7τ=0.7. 3. Semantic Consistency Score (SsemS_sem): We implement regex-based extraction to isolate experimental verbs (V) and numerical parameters (ℛPR). The component scores are defined as: Covstep=|gt∩gen||gt|,Eparam=|ℛgt∩ℛgen||ℛgt|.Cov_step= |V_gt _gen||V_gt|, E_param= |PR_gt _gen||PR_gt|. (9) To ensure numerical stability, if |gt|=0|V_gt|=0 or |ℛgt|=0|PR_gt|=0, the respective sub-score defaults to 0.5. The aggregated semantic score is computed as: Ssem=0.15⋅Covstep+0.15⋅Eparam+0.30⋅F1K+0.40⋅SR. splitS_sem&=0.15· Cov_step+0.15· E_param\\ & +0.30· F1_K+0.40· SR. split (10) Rationale for Weights: These coefficients were calibrated based on consultations with senior wet-lab biologists to prioritize structural completeness (SRSR) and keyword accuracy (F1KF1_K) over minor syntactic variations. This ensures the metric reflects practical laboratory utility rather than mere linguistic fluency. Scientific Validity (CsC_s). Since semantic overlap may miss subtle logical flaws, we leverage an LLM-as-a-Judge (Gemini-3-Flash) to audit protocols across five weighted dimensions: Critical Step Coverage (ScovS_cov), Parameter Precision (SparamS_param), Control Fidelity (SctrlS_ctrl), Biosafety (SsafeS_safe), and Objective Alignment (SobjS_obj). Cs=0.3Scov+0.2Sparam+0.15Sctrl+0.1Ssafe+0.25Sobj. splitC_s&=0.3S_cov+0.2S_param+0.15S_ctrl\\ & +0.1S_safe+0.25S_obj. split (11) E.2 Machine Code Generation Metrics Let SgtS_gt and SgenS_gen represent the sequences of executable operations. Code quality is evaluated through a three-layered approach: Rule-Based Validity. We strictly verify basic executability via deterministic checks: 1. Schema Compliance (SchemacompSchema_comp): The proportion of mandatory fields (e.g., node types, connection IDs) that adhere to the JSON definition. 2. Resource Validity (ResvalidRes_valid): The ratio of valid instrument IDs referenced in the code: Resvalid=1M∑i=1M[niid∈Ωvalid∪−1],Res_valid= 1M _i=1^MI[n_i^id∈ _valid∪\-1\], (12) where Ωvalid _valid represents the authorized hardware registry and −1-1 denotes manual steps. 3. Sequence Accuracy (AccseqAcc_seq): The Longest Common Subsequence (LCS) ratio measuring the ordering correctness of functional operations: Accseq=2⋅|LCS(Sgt,Sgen)||Sgt|+|Sgen|.Acc_seq= 2·|LCS(S_gt,S_gen)||S_gt|+|S_gen|. (13) Physical Compliance (CpC_p). A symbolic rule engine validates N hardware constraints (e.g., rpm≤15,000rpm≤ 15,000). The score penalizes Critical Errors (HALT, α=0.2α=0.2) and Warnings (WARN, β=0.05β=0.05): Cp=max(0,1−(αNhalt+βNwarn)).C_p= (0,1-(α N_halt+β N_warn) ). (14) This penalty mechanism ensures that safety violations significantly degrade the score, aligning with the zero-tolerance policy for hardware damage. Parameter Accuracy (AccparamAcc_param). Fine-grained parameter correctness is evaluated by LLM-as-a-Judge across Node Completeness (Sn-comS_n-com), Value Accuracy (Sp-accS_p-acc), and Field Completeness (Sp-comS_p-com): Accparam=0.3Sn-com+0.4Sp-acc+0.3Sp-com.Acc_param=0.3S_n-com+0.4S_p-acc+0.3S_p-com. (15) Overall Code Score (ScodeS_code). We employ a multiplicative formulation where SchemacompSchema_comp acts as a soft gating factor. Unlike additive metrics where accurate parameters might mask a broken syntax, this design ensures that if the foundational schema structure is flawed (Schemacomp→0Schema_comp→ 0), the overall executability score ScodeS_code is penalized towards zero. This reflects the operational reality that syntactically invalid code is non-executable regardless of semantic intent. Scode=Schemacomp⋅(0.35Accseq+0.25Accparam+0.30Cp+0.10Resvalid). split&S_code=Schema_comp· (0.35Acc_seq+\\ &0.25Acc_param+0.30C_p+0.10Res_valid ). split (16) E.3 System Performance Metrics To assess operational viability, we track: Efficiency: Execution Time (T), Total Steps (NstepN_step), and Token Consumption (NtokenN_token). Success Rate (Succ.Succ.): A task is deemed successful only if it meets strict domain validity criteria (Scode≥τcodeS_code≥ _code and Cp=1.0C_p=1.0): Succ.=∑i=1Ntotal[Taski is Valid]Ntotal.Succ.= _i=1^N_totalI[Task_i is Valid]N_total. (17) E.4 Statistical Analysis & Metric Verification To ensure the rigor of our evaluation, we conducted two types of analyses: (1) Cross-model verification to assess the objectivity of our LLM-as-a-Judge metric, and (2) Paired t-tests to determine the statistical significance of our performance gains. E.4.1 Metric Reliability (Cross-Judge) Table 11: Multi-Judge Consistency Analysis. Comparison of CsC_s scores across three distinct LLMs (N=640N=640 evaluations). GPT-4o confirms the trends with the highest correlation, validating the robustness of our metric. Model Variant Gemini Kimi GPT-4o BioProAgent-Full 0.591 0.597 0.515 w/o Knowledge 0.561 0.552 0.497 w/o Clarify 0.403 0.421 0.374 w/o FSM 0.262 0.285 0.105 Correlation (r) w/ Gemini 1.00 0.84 0.91 To rule out self-preference bias in the Gemini-based judge (CsC_s), we re-evaluated the ablation dataset using GPT-4o Islam and Moushi (2025) and Kimi-k2-Instruct Team et al. (2025). As shown in Table 11, the scoring trends across different reasoning engines are highly correlated (r≈0.91r≈ 0.91), confirming that the CsC_s metric objectively reflects protocol quality. GPT-4o exhibits a near-perfect correlation with Gemini (r=0.91r=0.91), validating the reliability of our primary metric. All three judges consistently rank BioProAgent-Full as the top performer and identify the removal of the FSM (w/o FSM) as the most detrimental ablation. Notably, GPT-4o assigns significantly lower scores to the w/o FSM variant (0.1050.105) compared to Gemini (0.2620.262), suggesting that advanced models penalize the lack of structured planning even more severely. E.4.2 Statistical Significance (P-values) Table 12: Overview of the Extended BioProBench. Benchmark Subset Scale & Granularity Evaluation Focus Key Metrics A: Scientific Drafting 160 Protocols Intent Reasoning • Intent Alignment (Cross-domain) (approx. 3k tokens/doc) Knowledge retrieval • Scientific Logic Score B: Automation Conversion 41 Sub-Experiments Local Code Mapping • Parameter Accuracy (Syn-Bio Focus) (22 Device APIs) Schema compliance • API Validity C: Long-Horizon Exec. 9 Protocols Global Orchestration • Global Success Rate (Full Pipeline) (547 Total Steps) State Persistence (Avg. 60.8 steps) • ID Consistency D: Error Correction 30 Injection Cases Deterministic Robustness • Detection Recall (Robustness) (Physical/Logic Errors) Self-healing capabilities • Fix Success Rate Table 13: Statistical Significance Analysis (Paired t-test). Comparison of BioProAgent against baselines and ablation variants (N=160N=160). All improvements are highly statistically significant. Category Comparison Pair (vs. Full Model) Mean Diff. P-value Baselines ReAct +0.136 <10−10<10^-10 AutoGPT +0.162 <10−11<10^-11 Reflexion +0.152 <10−11<10^-11 Ablations w/o Knowledge +0.030 0.025∗0.025^* w/o Clarify +0.188 <10−13<10^-13 w/o FSM +0.329 <10−49<10^-49 ∗ Significant (p<0.05p<0.05); All others are extremely significant (p<0.001p<0.001). We performed paired sample t-tests on the Scientific Validity scores (CsC_s) for the entire Subset A (N=160N=160). Table 13 reports the results for both the main baselines and the ablation variants. Appendix F Benchmark Details As shown in Table 12, we summary the extended BioProBench. Subset A (Protocol Drafting) 160 protocols across 15 biological domains with varying query richness to test intent disambiguation and scientific reasoning; Subset B (Code Generation) 41 synthetic biology sub-experiments with paired natural language drafts and ground-truth machine codes to evaluate hardware schema compliance and parameter accuracy; Subset C (Long-Horizon) tests 9 end-to-end long protocols, featuring up to 71 major steps, designed to test the limits of state persistence and context management; Subset D (Error Correction) measures 30 code snippets with injected physical/logic errors to evaluate the system’s deterministic self-healing capability. We enriched the environment with API schemas for 22 specialized synthetic biology devices and a comprehensive library of consumable identifiers. F.1 Statistical Significance of Subset C Table 14: Complexity Statistics of the 9 Long-Horizon Tasks in Subset C. Unlike standard benchmarks with short trajectories, these tasks involve up to 238 atomic steps and 169 unique consumables. A failure at any step leads to task failure. Task ID Device Interaction Nodes Atomic Operation Steps Involved Consumables Sample 1 15 55 31 Sample 2 4 24 6 Sample 3 8 16 17 Sample 4 17 35 26 Sample 5 18 55 42 Sample 6 19 27 31 Sample 7 71 238 130 Sample 8 36 81 169 Sample 9 8 16 38 Total 196 547 490 Average 21.8 60.8 54.4 While Subset C contains 9 high-level tasks, we argue that the sample size should be interpreted through the lens of trajectory complexity rather than task count. As detailed in Table 14, these are not single-turn queries but deep sequential decision processes. The tasks comprise a total of 547 atomic operation steps (avg. 60.8, max 238 per task). Executing Sample 7 requires managing 130 distinct consumables and 71 device interactions without a single hallucination. In wet-lab protocols, steps are strictly coupled. A deviation at Step 10 invalidates the result at Step 238. Therefore, BioProAgent’s 100% success rate in this subset represents 547 consecutive correct decisions in a grounded physical environment. If we assume a baseline agent has even a 95% step-wise accuracy, the probability of successfully completing the longest task (Sample 7, 238 steps) is negligible (0.95238≈5×10−60.95^238≈ 5× 10^-6). Our result demonstrates that the FSM-based architecture effectively prevents the compound error explosion typical in probabilistic LLMs. Table 15: Gallery of Autonomous Corrections. The table compares the raw design violation (left), the specific safety interlock triggered by the Rule Engine (center), and the rectified code after FSM-guided feedback (right). These examples correspond directly to the qualitative traces discussed in Section 4.5. Constraint Type Design Violation (sdesigns_design) Interlock Trigger (¬p _p) FSM Rectification (srectifys_rectify) Safety Alignment op.centrifuge( speed=25,000g, duration=15min) [HALT] Speed 25,000g exceeds rotor limit (Max: 15,000g). op.centrifuge( speed=15,000g, duration=15min) # Clamped to safe limit Resource Grounding op.transfer( source=new_plate, vol=50) [HALT] Resource ID new_plate not found in Registry Ω . op.transfer( source=plate_1, vol=50) # Mapped to valid grounded ID Causal Rectification op.seal_plate() op.add_reagent(vol=10) [HALT] Cannot add_reagent to a sealed container (Physical Collision Risk). op.add_reagent(vol=10) op.seal_plate() # Sequence causally reordered F.2 Subset D: Error Injection Taxonomy To further investigate the robustness of BioProAgent, we present a granular breakdown of performance metrics across three specific error categories in Subset D (Error Correction). Table 16 details the agent’s behavior when correcting Resource Grounding, Safety Alignment, and Causal Rectification. Table 16: Granular Performance Breakdown by Error Type (Subset D). While syntactic perfection remains challenging in complex repairs, BioProAgent maintains high CpC_p, reinforcing its role as a safety interlock. Error Type Counts CpC_p LCS Recall AccseqAcc_seq Resource Grounding 10 0.935 1.000 0.999 Safety Alignment 10 0.970 1.000 0.979 Causal Rectification 10 0.870 0.900 0.897 As shown in Table 16, BioProAgent demonstrates exceptionally high Physical Compliance (Cp=0.935C_p=0.935) and flawless LCS Recall (1.000) for resource hallucinations. This indicates that while the agent correctly identifies the semantic intent and enforces safety limits, strictly formatting the JSON schema during complex repairs remains a non-trivial challenge. Crucially, the high CpC_p scores confirm that the neuro-symbolic FSM successfully prevents dangerous instructions even when code syntax is imperfect. F.3 The Error Correction Gallery (Subset D) Table 15 details specific instances of autonomous rectification captured during the ablation study (corresponding to the taxonomy in Appendix F.2). These examples illustrate the three primary categories of hard constraints enforced by our Physical Rule Engine (pK_p), demonstrating how the FSM deterministically intercepts hazardous instructions before physical execution. Appendix G Impact of Scaling to “Thinking” Models Table 17: Evaluation of “Thinking” Reasoning Models on Subsets C and D. Subset C Subset D Method Backbone Succ. ↑ AccparamAcc_param ↑ Corr. Succ. ↑ CpC_p ↑ Time (s) ↓ Direct DeepSeek-V3.2-Think 100.0% 0.414 93.3% 0.948 196.2 Direct Gemini-3-Pro-Think 100.0% 0.439 73.3% 0.765 81.4 BioProAgent DeepSeek-V3.2-Think 55.6% 0.366 93.3% 0.955 1313.3 BioProAgent Gemini-3-Pro-Think 44.4% 0.256 90.0% 0.922 316.1 BioProAgent Gemini-3-Flash 100.0% 0.718 90.0% 0.925 128.0 A critical question regarding the “Execution Gap” is whether scaling up to massive, state-of-the-art “System 2” reasoning models (e.g., DeepSeek-V3.2-Think Liu et al. (2025a), Gemini-3-Pro-Think Google and Google DeepMind (2025)) can inherently solve the alignment and error correction challenges, rendering neuro-symbolic frameworks unnecessary. To investigate this, we evaluated these models on Subsets C and D, both via direct prompting and as the cognitive engine inside BioProAgent. The Capability Plateau. As shown in Table 17, massive reasoning models exhibit strong single-shot robustness due to their extensive internal <think> trajectories. Direct DeepSeek-V3.2-Think achieves a 93.3% correction rate on its own. However, relying solely on massive models scales poorly for iterative robotic tasks and fails to capture the fine-grained sequence logic in long-horizon tasks (achieving only Accparam=0.414Acc_param=0.414 on Subset C). Reasoning Context Overflow. Counter-intuitively, when integrating these heavy reasoning models inside BioProAgent’s iterative loop, performance on long-horizon tasks degraded (Subset C Success Rate dropped to 44%-55%). BioProAgent’s FSM acts as an external multi-step framework. When models generate thousands of tokens of internal <think> traces at every FSM state transition, the working memory rapidly saturates (averaging >66,000>66,000 tokens). This causes “Attention Dilution,” leading the model to forget strict JSON formatting constraints and fail prematurely. The Efficiency Triumph. The true value of our neuro-symbolic architecture lies in democratizing high-fidelity autonomy. By providing a deterministic external reasoning track, BioProAgent empowers a highly efficient, lightweight model (Gemini-3-Flash) to achieve a 90.0% correction rate (Cp=0.925C_p=0.925). Crucially, BioProAgent with Flash resolves physical violations in an average of 128.0 seconds, whereas BioProAgent with DeepSeek-V3.2-Think requires 1313.3 seconds. This demonstrates that for strict physical execution, combining an agile instruction-following model with a rigid symbolic controller is a far more pragmatic solution than blindly scaling reasoning compute. Table 18: Side-by-Side Execution Log Comparison. BioProAgent utilizes FSM states to progress deterministically, completing the task in 8 steps. In contrast, AutoGPT enters a semantic drift loop, exhausting the context window (Step 45) without producing valid code. Step BioProAgent (Ours) AutoGPT (Baseline) 1 [State: CLARIFY_INTENT] Trigger: σambiguous _ambiguous Action: clarify_experiment_scope(query=...) Output: "Sample Volume: 50uL, Kit ID: 10038" [Thought] I need to find the protocol for metabolomics. Action: retrieve_knowledge(query="metabolomics prep") 2 [State: DESIGN_DRAFT] Trigger: σknow∧¬σdraft _know _draft Action: generate_scientific_draft(...) Grounding (Φ ): Mapped "Kit_A" → ID:10038 [Thought] I found some docs. Now I need specific reagent list. Action: retrieve_knowledge(query="reagent list") 3 [State: VERIFY_DRAFT] Trigger: σdraft∧(σsci=0) _draft ( _sci=0) Action: reflect_on_protocol(...) Result: Scientific Logic Approved (σsci←1 _sci← 1). [Thought] I need to check the temperature settings. Action: retrieve_knowledge(query="temperature settings") … (Step 4-7: State transitions through DESIGN_CODE and Physical Rule Verification pK_p) (Repeated retrieval loops for 30+ steps…) 8 / 45 [State: SUCCESS] Trigger: σcode∧(σphys=1) _code ( _phys=1) Result: Valid JSON Code committed. Task Success (Total Tokens: 10,329) [Error] Max Context Exceeded. Result: Timeout / Fail. Task Failure (Total Tokens: 59,278) Appendix H Case Studies & Logs H.1 Detailed Execution Traces (Subset C) To provide granular insight into the “Efficiency Gap” visualized in Figure 5, we present raw execution logs from a representative long-horizon task (Task ID: Metabolomics-Sample-Prep). Table 18 contrasts the decision-making trajectory of BioProAgent against AutoGPT. While BioProAgent leverages its deterministic FSM backbone (st+1=Δ(σt,st)s_t+1= ( _t,s_t)) to strictly enforce the Design-Verify-Rectify cadence, AutoGPT succumbs to a “semantic drift loop,” repeatedly querying redundant information without converging on an executable plan. H.2 Taxonomy of Baseline Failures Expanding on the trace analysis in Figure 5, we categorize the failure modes of baseline agents into two distinct pathologies based on the execution logs across our benchmark evaluations: Context Paralysis (Dominant in ReAct). As observed in the vast majority of ReAct failures (visualized in Figure 5, Middle), the agent correctly retrieves the schema but fails to synthesize it into code. The logs show abrupt termination with empty outputs. This confirms that without Symbol Grounding (Φ ), the raw HTML/JSON payloads from API docs quickly saturate the LLM’s working memory, leading to a cognitive freeze. Info-Seeking Loop (Dominant in AutoGPT). In almost all instances of AutoGPT failures (visualized in Figure 5, Bottom), the agent enters a recursive retrieval loop (e.g., querying buffer composition → “buffer pH” → “buffer supplier”). Unlike BioProAgent’s deterministic FSM, which mandates a rigid progression from DESIGN_DRAFT through VERIFY_DRAFT to DESIGN_CODE, AutoGPT lacks a structural stopping condition, optimizing for information completeness rather than task completion. Appendix I System Prompts To support reproducibility, we provide the core system prompts used in BioProAgent. Variable placeholders are denoted by brackets (e.g., user_input). I.1 FSM-Driven Planner Prompt The planner prompt illustrates how the Finite State Machine (FSM) explicitly governs the agent’s trajectory. Note the State->Action Mapping section, which enforces the Priority Decision Matrix described in Section 3.2 (Table 1), ensuring that the agent strictly adheres to the Design-Verify-Rectify workflow. ⬇ You are BioProAgent, a bio-experiment planning agent. Your task is to convert user requests into executable tool call sequences. ## Scientific Principles - Reproducibility: Steps are clear and reproducible. - Control Principle: Negative control (NTC), Positive control (PTC), Blank control. - Variable Control: Change only one variable at a time. ## Current State: current_state ## Suggested Actions: state_guidance ## User Request user_input ## Experiment Context experiment_context ## Executed Steps execution_summary ## Available Tools tools_description ## Output Format Requirements (Strictly Follow) Output must be a JSON array. Example: ["tool_name": "retrieve_knowledge", "args": ...] ### State -> Action Mapping (Hard Constraints) - CLARIFY_INTENT -> [clarify_experiment_scope] - DESIGN_DRAFT -> [retrieve_knowledge, generate_scientific_draft] - VERIFY_DRAFT -> [reflect_on_protocol] - RECTIFY_DRAFT -> [modify_protocol, reflect_on_protocol] - DESIGN_CODE -> [align_draft_to_automation, generate_machine_code] - RECTIFY_CODE -> [fix_machine_code, validate_machine_code] - SUCCESS -> [ask_user_confirmation] ## Important Instructions - Follow scientific principles strictly. - Current state is **current_state** - Suggested actions: state_guidance - The FSM state determines your valid action space. Do not hallucinate tools outside the mapping. Listing 1: The Neuro-Symbolic Planner Prompt I.2 Neuro-Symbolic Alignment Prompt This prompt operationalizes the Semantic Symbol Grounding (Φ ) phase. It maps natural language protocols to the hardware registry by categorizing steps into [AUTO], [EXTERNAL], and [MANUAL], and strictly grounding natural language reagents to valid registry IDs to prevent context overflow. ⬇ You are a biological experiment automation expert. Task: Convert the [Reference Info] into a standardized executable process. ## Step Tagging Rules (Critical!) Every step must be tagged with one of the following: | Tag | Meaning | Usage | |-----|---------|-------| | **[AUTO]** | Executable by platform | Use available devices below | | **[EXTERNAL]** | Requires external device | Lab has device, but not integrated | | **[MANUAL]** | Requires human operation | Decision making or special ops | **Important**: [EXTERNAL] and [MANUAL] are NOT failures. They are normal parts of complex biological workflows. ## Symbol Grounding ($ $) Rules When assigning an [AUTO] step, you MUST map fuzzy reagent/container names to strict semantic pointers (e.g., map "Kit_A" -> "ID:10038") based on the Provided Context. Do NOT pass full payload schemas into the sequence. ## Available Automation Devices (Summary) - Pipetting Workstation: Liquid transfer (0.5uL-1mL), mixing, temp control - PCR Machine: Temp cycling (4-99C) - Sealer/Peeler: Plate sealing - Centrifuge: Separation (500-15000g) - Incubator: Constant temp (20-60C) - Plate Reader: Absorbance/Fluorescence ## Output Format Specification 1. [AUTO] Pipetting Workstation - Setup PCR Reaction 1.1 [AUTO] Pipetting Workstation - Add Template DNA (Vol=2uL) - Source: [Grounded_ID: template_tube_1] - Target: [Grounded_ID: pcr_plate_01] 1.2 [AUTO] Pipetting Workstation - Add Master Mix 2. [AUTO] Sealer - Seal Plate (Temp=170C) 3. [AUTO] PCR Machine - Amplification - Program: 95C 5min -> [95C 30s, 58C 30s, 72C 1min] x35 4. [EXTERNAL] Gel Electrophoresis - Check Products - Note: Requires external imaging system 5. [MANUAL] Human Decision - Verify Band Size - Criteria: Target band at ~500bp ## Input Information User Requirements: exp_info Reference Protocol: protocol Listing 2: Protocol Alignment and Device Grounding Prompt I.3 Baseline SOP Injection To ensure fair comparison, baselines (ReAct, Direct LLM) were injected with a Standard Operating Procedure (SOP) prompt. This minimizes naive formatting errors and ensures that performance gaps are attributable to reasoning architecture rather than prompt engineering. ⬇ ## Laboratory Standard Operating Procedures (SOP) You are a Senior Laboratory Automation Engineer. To ensure success, you MUST follow these protocols: 1. **Tool Priority**: Do NOT hallucinate JSON formats. - Use ‘generate_machine_code‘ to create the final JSON. - This tool knows the correct Schema (nodes, resourceId, connections). 2. **Validation Loop**: - After generating code, you MUST use ‘validate_machine_code‘. - If validation fails (red light), use ‘fix_machine_code‘ or regenerate. - Do NOT output the final answer until validation passes. 3. **Data Handling**: - Use data references (e.g., $draft, $machine_code) to pass information between tools to avoid context overflow. Listing 3: Standard Operating Procedure (SOP) for Baselines