Paper deep dive

Evaluator-in-the-Loop Monte Carlo Tree Search via LLM Agents for Motif Scaffolding in Protein Design

Haotian Hu, Oguzhan Gungordo, Siheng Xiong, Faramarz Fekri

Year: 2026Venue: arXiv preprintArea: cs.MAType: PreprintEmbeddings: 47

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 10/5/2026, 4:57:02 AM

Summary

The paper introduces ELMS (Evidence-based LLM-guided Monte Carlo Search), a framework for motif scaffolding in protein design that integrates Large Language Model agents into a Monte Carlo Tree Search (MCTS) loop. Unlike traditional generate-then-filter methods, ELMS uses structural evaluation feedback to diagnose failures via a Critic Agent and select targeted sequence modifications via a Policy Agent, while motif-locked operators ensure structural validity. This approach significantly outperforms baselines on GeomMotif and MotifBench benchmarks.

Entities (9)

ELMS · method · 98%MCTS · algorithm · 95%Critic Agent · component · 92%Policy Agent · component · 92%GeomMotif · benchmark · 90%MotifBench · benchmark · 90%ESMFold · tool · 85%ProteinMPNN · tool · 85%gpt-5-nano · model · 80%

Relation Signals (10)

ELMS → evaluatedon → MotifBench

confidence 95% · On MotifBench... it solves 26.7 of 30 tasks on average

ELMS → evaluatedon → GeomMotif

confidence 95% · Under the standard GeomMotif protocol... ELMS achieves Successful rates of 86.41%

ELMS → outperforms → strongest prior baseline

confidence 95% · exceeding the strongest prior baseline by 19.3 and 21.9 percentage points, respectively.

ELMS → uses → MCTS

confidence 95% · ELMS therefore retains evaluated scaffolds as persistent search states... and MCTS determines which historical states should receive further design effort.

ELMS → contains → Policy Agent

confidence 92% · a Policy Agent selects targeted operators with execution parameters

ELMS → contains → Critic Agent

confidence 92% · a Critic Agent diagnoses state-local structural failures

Critic Agent → diagnoses → structural failures

confidence 90% · a Critic Agent diagnoses state-local structural failures

Policy Agent → selects → Operators

confidence 90% · a Policy Agent selects targeted operators with execution parameters

Critic Agent → implementedwith → gpt-5-nano

confidence 85% · Both agents use gpt-5-nano with constrained JSON outputs

Policy Agent → implementedwith → gpt-5-nano

confidence 85% · Both agents use gpt-5-nano with constrained JSON outputs

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Motif-scaffolding systems commonly follow a generate-then-filter paradigm, in which candidate proteins are generated independently and structural evaluation is used primarily for terminal screening or ranking. This paradigm underuses evaluation: failed predictions contain state-specific evidence about whether a design requires repair of motif geometry, global foldability, or other structural constraints. We introduce \textbf{ELMS} (Evidence-based LLM-guided Monte Carlo Search), an evaluator-in-the-loop search framework for motif scaffolding that turns such evaluator feedback into targeted design actions. Effective reuse of structural feedback is nontrivial because different scaffold states exhibit different failure modes, and repeatedly refining a single trajectory can prematurely commit computation to an unproductive region of sequence space. ELMS therefore retains evaluated scaffolds as persistent search states: a Critic Agent diagnoses state-local structural failures, a Policy Agent selects targeted operators with execution parameters, motif-locked operators realize legal sequence modifications, and MCTS determines which historical states should receive further design effort. Under the standard GeomMotif protocol (100 candidates per task), ELMS achieves Successful rates of 86.41\% on single-motif tasks and 84.57\% on paired-motif tasks, exceeding the strongest prior baseline by 19.3 and 21.9 percentage points, respectively. On MotifBench, under a matched 100-candidate search budget, it solves 26.7 of 30 tasks on average (88.89\% Task Success), compared with 16.0 tasks (53.33\%) for the strongest baseline. These results establish ELMS as an effective approach for converting structural evaluation from a terminal filter into actionable guidance for iterative motif scaffolding.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

44,102 characters extracted from source content.

Expand or collapse full text

Evaluator-in-the-Loop Monte Carlo Tree Search via LLM Agents for Motif Scaffolding in Protein Design Haotian Hu1 Oguzhan Gungordo1 Siheng Xiong1 Faramarz Fekri1 1Georgia Institute of Technology Abstract Motif-scaffolding systems commonly follow a generate-then-filter paradigm, in which candidate proteins are generated independently and structural evaluation is used primarily for terminal screening or ranking. This paradigm underuses evaluation: failed predictions contain state-specific evidence about whether a design requires repair of motif geometry, global foldability, or other structural constraints. We introduce ELMS (Evidence-based LLM-guided Monte Carlo Search), an evaluator-in-the-loop search framework for motif scaffolding that turns such evaluator feedback into targeted design actions. Effective reuse of structural feedback is nontrivial because different scaffold states exhibit different failure modes, and repeatedly refining a single trajectory can prematurely commit computation to an unproductive region of sequence space. ELMS therefore retains evaluated scaffolds as persistent search states: a Critic Agent diagnoses state-local structural failures, a Policy Agent selects targeted operators with execution parameters, motif-locked operators realize legal sequence modifications, and MCTS determines which historical states should receive further design effort. Under the standard GeomMotif protocol (100 candidates per task), ELMS achieves Successful rates of 86.41% on single-motif tasks and 84.57% on paired-motif tasks, exceeding the strongest prior baseline by 19.3 and 21.9 percentage points, respectively. On MotifBench, under a matched 100-candidate search budget, it solves 26.7 of 30 tasks on average (88.89% Task Success), compared with 16.0 tasks (53.33%) for the strongest baseline. These results establish ELMS as an effective approach for converting structural evaluation from a terminal filter into actionable guidance for iterative motif scaffolding. 1 Introduction Motif scaffolding connects a local hypothesis of molecular function to a physically realizable protein. A small set of residues may specify catalysis, binding, or coordination through a precise three-dimensional arrangement, as demonstrated in computational enzyme and epitope scaffolding (Correia et al., 2014). The surrounding protein, however, must stabilize this local geometry while forming a globally foldable and structurally coherent scaffold, motivating structure-conditioned scaffold design approaches such as constrained hallucination and inpainting (Wang et al., 2022). More recently, diffusion-based protein design has enabled direct generation of diverse scaffolds around prescribed structural motifs (Watson et al., 2023; Butcher et al., 2025), with subsequent methods extending this paradigm toward more flexible and atomically specified motif constraints (Ahern et al., 2026). Despite these advances, motif scaffolding still couples several objectives that need not improve together: preserving the prescribed motif geometry, maintaining global foldability, and producing structures sufficiently distinct from known or reference proteins. Reliably balancing these local and global requirements therefore remains a central challenge in functional protein design. Most modern methods address motif scaffolding through a generate-then-filter paradigm. A motif-conditioned generator first produces a population of candidate scaffolds, after which structure prediction and benchmark evaluation are used to accept, reject, or rank the resulting designs. Diffusion-based approaches such as RFdiffusion have made motif-conditioned backbone generation substantially more effective (Watson et al., 2023), while flow-matching methods have provided an alternative framework for sampling conditional protein structures (Yim et al., 2024). Subsequent methods such as Genie 2 Lin et al. (2024) and FADiff Liu et al. (2024) have expanded the range and flexibility of geometric constraints that can be imposed during generation. More recent approaches extend motif-conditioned design toward finer-grained and atomistic specifications (Ahern et al., 2026; Geffner et al., 2026). Despite these advances in generation, structural evaluation is still typically consumed only after candidates have been produced. Low structural confidence and motif displacement indicate different failure modes and therefore suggest different corrective actions, yet under a generate-then-filter interface they may both be treated simply as unsuccessful samples. Consequently, evaluation is primarily used to measure and filter failures rather than to determine how the next design should be changed. GeomMotif Strashnov et al. (2026) and MotifBench Zheng et al. (2025) evaluate motif-scaffolding methods under benchmark-specified candidate populations, measuring whether generated designs satisfy success criteria and, in some analyses, novelty or diversity. Our question is complementary: when structural evaluation, high-accuracy simulation, or expert review is costly, can evaluator feedback be used during inference to discover at least one feasible scaffold within a fixed candidate/evaluator budget more reliably than generate-then-filter or simpler feedback-guided search? To address this question, we introduce Evidence-based LLM-guided Monte Carlo Search. ELMS treats evaluated scaffolds as persistent search states: a Critic Agent summarizes state-local structural failures, a Policy Agent selects a legal operator with execution parameters, motif-locked operators realize the edit, and MCTS decides which states receive further design effort. This turns structural evaluation from a terminal filter into an inference-time signal for allocating a limited design budget. Figure 1: Overview of ELMS. The search tree selects an evaluated state for expansion; the Critic Agent analyzes the state-local evidence to identify which structural or task-specific constraints are limiting the current candidate, the Policy Agent translates this critic into a concrete edit plan over mutable sequence regions, and a motif-locked operator applies the edit while preserving the required motif positions to generate a new candidate sequence. Structural evaluation returns the resulting child to the search tree, where value and visit information is backpropagated to guide subsequent expansions. After search, evaluated states are ranked deterministically for benchmark reporting. Our main contributions are: • We formulate motif scaffolding as ELMS, which conducts Monte Carlo Tree Search with an evaluator in the loop. We therefore explicitly separate internal candidate-generation/evaluation budget from the final benchmark reporting interface. • We introduce a multi-agent MCTS controller in which both LLM Critic and Policy Agents operate over state-local structural evidence, while motif-locked operators enforce constraints. • We evaluate the framework on two complementary motif-scaffolding benchmarks. Under the standard GeomMotif Success protocol, it achieves Successful rates of 86.41% on single-motif tasks and 84.57% on paired-motif tasks. In a controlled MotifBench setting with 100 evaluated candidates per case, our framework solves 26.7 of 30 tasks on average, corresponding to 88.89% Task Success. We further use matched-budget ablations and budget-scaling experiments to isolate the contribution of evaluator-guided tree search and analyze its sensitivity to inference compute. 2 Related Work Motif-conditioned protein generation. Computational motif scaffolding has progressed from physics-based scaffold search to learned generative models capable of constructing proteins directly around prescribed structural motifs. Classical Rosetta-based approaches placed functional residues into existing or newly constructed scaffolds using energy-based optimization and conformational sampling (Kuhlman et al., 2003; Röthlisberger et al., 2008; Jiang et al., 2008; Correia et al., 2014). Deep-learning approaches subsequently removed the need for predefined scaffold libraries: constrained hallucination and RoseTTAFold inpainting generated structures around functional sites (Wang et al., 2022), while RFdiffusion established diffusion-based conditional backbone generation as a strong motif-scaffolding paradigm (Watson et al., 2023). Later methods expanded the flexibility of motif conditioning through SE(3) flow matching, sequence diffusion, and multi-motif generation, including FrameFlowYim et al. (2024), DPLMWang et al. (2024), Genie 2Lin et al. (2024), and FADiff (Liu et al., 2024). More recently, RFdiffusion2 and La-Proteina have extended motif conditioning toward unindexed and atom-level specifications (Ahern et al., 2026; Geffner et al., 2026). Standardized benchmarks such as MotifBenchZheng et al. (2025) and GeomMotifStrashnov et al. (2026) expose complementary limitations of these models: MotifBench evaluates function-derived catalytic and binding motifs, whereas GeomMotif isolates arbitrary geometric preservation across diverse structural contexts. Evaluation-guided and agentic protein design. Modern protein-design pipelines commonly combine inverse folding and structure prediction, using ProteinMPNN, AlphaFold, or ESMFold to test whether generated sequences recover the intended structures (Dauparas et al., 2022; Jumper et al., 2021; Lin et al., 2023; Hayes et al., 2025). Most motif-scaffolding pipelines use these evaluations primarily for filtering, ranking, or benchmark scoring. Related work has begun to use external feedback more actively: guidance-based methods such as ProteinGuide steer pretrained protein generators toward desired properties, while agentic systems such as ProtAgents coordinate language models, scientific knowledge, and computational tools during protein-design workflows (Xiong et al., 2026a; Ghafarollahi and Buehler, 2024). These approaches demonstrate the value of incorporating additional evidence into design, but generally treat feedback as a global conditioning signal or tool-level reasoning resource rather than as candidate-specific evidence for deciding how an evaluated scaffold should be edited next. Search and planning with evaluator feedback. Tree search methods provide a general mechanism for allocating inference-time computation toward promising branches while preserving exploration (Kocsis and Szepesvári, 2006; Browne et al., 2012; Xiong et al., 2026b; Gungordu et al., 2026; Xiong et al., 2026c). In motif scaffolding, fold confidence, motif geometry, and hard validity expose different online failure modes, while novelty and redundancy are most naturally evaluated post hoc on the accepted candidate set. Our work uses ELMS as a decision interface around fixed structural evaluators. Unlike generate-then-filter pipelines, each evaluated scaffold becomes a state whose evidence can guide the next expansion; unlike unconstrained agentic design, LLM outputs are limited to critic summaries and operator selection with execution parameters, while non-LLM modules enforce motif locks, realize child sequences, evaluate structures, and back propagate values through the tree. 3 Preliminaries 3.1 Evaluation of Success and Novelty We follow the evaluation protocol of GeomMotif (Strashnov et al., 2026) and MotifBench Zheng et al. (2025) to assess whether a generated scaffold both preserves the target motif geometry and departs from the reference fold. Success. We use benchmark-specific success definitions. In GeomMotif, we use the benchmark’s modality-specific success criteria. For our sequence-generated candidates, success requires motif-residue CαC_α RMSD ≤1.0Å≤ 1.0\, and mean pLDDT ≥70≥ 70. Structure-generator baselines retain the benchmark’s native structure-level evaluation protocol, which uses the motif CαC_α RMSD gate together with a structure-level self-consistency gate rather than pLDDT. In MotifBench, success is defined at the submitted backbone level: each submitted backbone is sequence-designed eight times with ProteinMPNN and refolded with ESMFold, and the backbone is successful if at least one refolded design satisfies both motif RMSD <1.0Å<1.0~ over motif-residue N, CαC_α, and C atoms and backbone self-consistency RMSD <2.0Å<2.0~ over backbone CαC_α atoms. Following MotifBench, no pLDDT or confidence filter is applied to the final backbone-success decision. Novelty. Success alone does not distinguish genuinely new scaffolds from candidates that simply reconstruct the original reference fold. Therefore, novelty is evaluated among successful candidates by comparing each generated scaffold to the corresponding ground-truth structure using TM-score. We denote the ground-truth reference scaffold for the current task as SrefS_ref. A successful candidate is treated as novel when its TM-score to the reference structure is below 0.80.8: Novel(St)=[Success(St)=1∧TMscore(St,Sref)≤0.8].Novel(S_t)=I [Success(S_t)=1\; \;TMscore(S_t,S_ref)≤ 0.8 ]. 3.2 Action Library and Motif-Locked Operators We use a fixed LLM-synthesized library of biologically motivated motif-locked operators, distilled from protein-design literature, frozen before evaluation, and reused unchanged across GeomMotif and MotifBench. During search, the Policy Agent selects an operator identifier and execution parameters. Actions are realized through fixed-length masked decoding with a protein language model. Decoding may be stochastic, but modifies only permitted scaffold positions while preserving prescribed motif residue identities and their assigned sequence positions. Folded motif geometry is assessed separately by the structure evaluator. The operator inventory, routing, and decoding details are provided in Appendix. 4 Methodology 4.1 Framework Overview ELMS improves motif scaffolding by using LLM agents to transform state-local structural feedback into targeted intervention decisions, enforcing motif preservation through motif-locked sequence operators, and adaptively allocating search effort toward promising or repairable design trajectories. We formulate the search as a bounded MCTS problem (Kocsis and Szepesvári, 2006; Browne et al., 2012). Each tree node stores an evaluated motif-constrained scaffold state Si=(xi,Πi,ℓi,pi,aiin,Ei,Ni,Wi,Qi).S_i=(x_i, _i, _i,p_i,a_i^in,E_i,N_i,W_i,Q_i). Here, i indexes a node in the MCTS tree. xi∈Lix_i ^L_i is the protein sequence over amino-acid alphabet A, and LiL_i is its total sequence length. ℓi _i records the scaffold lengths around the motif segments: if a task contains m ordered motif segments, then ℓi=(ℓi,0,ℓi,1,…,ℓi,m) _i=( _i,0, _i,1,…, _i,m) records the length before the first motif segment, the lengths between adjacent motif segments, and the length after the last motif segment. Together with the fixed motif segment lengths, these values determine LiL_i. Πi _i is the motif-to-scaffold placement map, which maps each motif residue to a position in 1,…,Li\1,…,L_i\ while satisfying benchmark placement constraints and preserving motif order. The parent pointer pip_i denotes the parent node of SiS_i, and aiina_i^in records the incoming motif-locked action specification that generated SiS_i from that parent. EiE_i is the cached structural evidence obtained when SiS_i is created, including motif RMSD, pLDDT, hard-validity signals, and online feasibility-proxy margins. (Ni,Wi,Qi)(N_i,W_i,Q_i) are the visit count, accumulated value, and mean value used by MCTS. As shown in Fig. 1, each iteration follows a select–critique–choose-action–apply–evaluate–backpropagate loop. Here, superscript (r)(r) indexes the current expansion iteration and is separate from the persistent node index i. The selected parent state is Spar(r)S_par^(r), and the newly generated child state is Schild(r)S_child^(r). The Critic Agent reads the cached evidence Epar(r)E_par^(r) associated with Spar(r)S_par^(r) and produces a critique C(r)C^(r). Conditioned on (Spar(r),Epar(r),C(r))(S_par^(r),E_par^(r),C^(r)), the Policy Agent returns a legal action specification a(r)a^(r) from the operator library in Appendix. This specification includes an operator identifier and execution parameters such as search scale, mutable positions, repair fraction, and sampling temperature. The corresponding motif-locked operator applies the action to legal scaffold positions while preserving the prescribed motif residues, producing Schild(r)S_child^(r). The child state is then folded and evaluated to obtain Echild(r)E_child^(r), and the resulting search utility is backpropagated to update the MCTS statistics. MCTSTreeMCTS\ TreeSpar(r),Epar(r)S_par^(r),E_par^(r)C(r)C^(r)a(r)a^(r)Schild(r)S_child^(r)Echild(r)E_child^(r)selectCritiqueSelect actionApplyEvaluatebackpropagation 4.2 Search and Submission Budgets Let BsearchB_search denote the internal search budget, defined as the number of candidate states evaluated by the structural prediction and evaluation pipeline. Let KsubK_sub denote the number of final designs retained for benchmark reporting. These two quantities correspond to different practical stages: the search procedure may evaluate multiple intermediate states while exploring and repairing design trajectories, whereas a benchmark or downstream workflow may retain only a smaller reported set. We use this distinction to separate internal search effort from the final reporting interface when analyzing budget sensitivity. When the terminal ranking uses the benchmark success indicator, task solving at retained size KsubK_sub is interpreted as whether the evaluated pool contains a successful candidate under BsearchB_search, not as an independent blind Top-KsubK_sub selection problem. 4.3 LLM Agents and Non-LLM Modules Table 1 separates the two LLM components, the Critic and Policy Agents, from non-LLM code and fixed pretrained protein-model inference. Both agents use gpt-5-nano with constrained JSON outputs; all numeric measurements, constraint checks, tree statistics, sequence edits, refolding calls, and benchmark metrics are produced outside the LLM. Table 1: Role map for ELMS. The table identifies which roles are dedicated LLM agents and which ones are non-LLM modules or fixed protein-model calls. Role Implementation LLM? Output MCTS search loop Deterministic UCB/back propagate functions No Selected tree node, visit counts, back propagate values Critic Agent gpt-5-nano with ontology-constrained JSON Yes Candidate-local critic labels and rationales Policy Agent gpt-5-nano with action-schema-constrained JSON Yes Legal operator choices plus execution parameters Motif-locked operator Motif-lock template plus sequence-model backend No Apply selected actions to legal scaffold regions while preserving motif residues Structure evaluator ESMFold plus benchmark-specific adapters No Fold confidence, motif RMSD, hard-validity signals, online proxy status Benchmark reporter Deterministic ranking procedure No Ranked evaluated candidate set This design prevents the LLM from inventing benchmark scores or bypassing motif constraints. The agents produce only constrained critic labels and legal action-schema outputs, including operator identifiers and execution parameters; motif-locked operators then apply the selected edits only to legal scaffold regions while preserving prescribed motif residues; motif-lock and validity checks are enforced deterministically. All numeric measurements and benchmark metrics come from the structure evaluator, not the LLM. The no-leakage interfaces, prompts, and schema validation rules are provided in Appendix. 4.4 Task Representation and Blinded State Initialization A benchmark instance is represented as =(M,C,L,Ref),T=(M,C,L,Ref), where M specifies public motif residues and fixed identities, C contains placement or contig constraints, L is the allowed length range and RefRef contains benchmark-provided reference information used only by the evaluator. The framework sees only the public motif mask, public motif identities, generated sequences, predicted structures, and candidate-local structural feedback. For both GeomMotif and MotifBench, each task is initialized from benchmark-provided motif constraints rather than from a benchmark-origin scaffold sequence. The task adapter reads the required motif specification, including motif residues or segments, target length or allowed length range, motif-placement information, and reference motif coordinates. The initial state S0S_0 is a virtual motif-only root that stores these sanitized constraints but is not itself structurally evaluated. Root expansion uses legal restart or budgeted sequence-only operators to sample complete scaffold sequences by masking all non-motif scaffold positions under the chosen length and placement constraints. These operators preserve the prescribed motif residues, modify only legal scaffold regions, and provide broad sequence initializations before any evaluated parent exists. After the first complete candidates are generated, they become the first evaluated tree nodes. All subsequent states are produced by MCTS expansion, motif-locked sequence modification, and fresh structural evaluation under the corresponding benchmark adapter. 4.5 MCTS Selection, Expansion, Evaluation, and Backpropagation At iteration r, the search procedure descends the current tree using an upper-confidence criterion. For an already evaluated child node j of parent node i with Nj>0N_j>0, the default selection score is UCB⁡(i,j)=Qj+clog⁡(Ni+1)Nj+1.UCB(i,j)=Q_j+c (N_i+1)N_j+1. Here, NiN_i is the number of backed-up child evaluations whose selected path contains node i, and NjN_j is the corresponding count for child node j. Thus, if a batch expansion evaluates multiple children with a shared ancestor, that ancestor receives one visit update per evaluated child. QjQ_j is defined only after Nj>0N_j>0 and equals the accumulated backpropagated value divided by NjN_j. Legal but unevaluated children are expanded before this score is computed; the +1+1 terms stabilize exploration among visited nodes and do not define QjQ_j for an unvisited child. The exploration constant c controls how strongly the search favors less-visited children. Among visited children, deterministic tie-breaking is performed by depth, hard-validity, and evaluator margin before repeatedly exploiting the current best-valued branch. Expansion converts an evaluated leaf into one or more legal action specifications. The Critic Agent receives a compact node record containing the sequence length, public motif mask, parent action, current online metrics, evidence records, and visit statistics. It emits ontology labels such as quality_failed, motif_geometry_failed, near_threshold, or state_successful. The Policy Agent then maps the verified critic to legal operator choices and execution parameters from the operator library. Each accepted action is realized by the motif-locked operator backend rather than by the LLM. The backend constructs an editable mask over non-motif residues, preserves motif identities and motif-to-sequence mapping, validates the amino-acid alphabet and length constraints, removes duplicates, and only then admits the child for evaluation. Newly generated sequences are folded with ESMFold (Lin et al., 2023); benchmark adapters compute fold confidence, motif RMSD, online feasibility-proxy margins, and hard-constraint flags. Final benchmark success, including MotifBench backbone-level success, is recomputed separately under the benchmark protocol. After evaluation, the newly evaluated child node j receives a scalar search utility RjR_j, computed from its state-local evaluation record as detailed in Appendix. This value is then backpropagated to every ancestor node i on the selected path: Ni←Ni+1,Wi←Wi+Rj,Qi←Wi/Ni.N_i← N_i+1, W_i← W_i+R_j, Q_i← W_i/N_i. The backpropagated value affects future tree selection only. The individual metrics remain available to the Critic Agent, Policy Agent, and terminal selector so that high-confidence but geometrically incorrect candidates are not collapsed into the same explanation as low-confidence but geometry-preserving candidates. 4.6 Termination and Benchmark-Level Reporting Iteration in Fig. 1 stops when the tree reaches the evaluator budget or when no legal children remain. Benchmark reporting uses only structurally evaluated states. For each task t, let tV_t denote the set of evaluated states in the MCTS tree. We rank states in tV_t by a fixed deterministic ordering: hard-valid candidates are preferred first, followed by benchmark success, lower motif RMSD and higher fold confidence. When a benchmark requires a retained set, tKsub⊆tP_t^K_sub _t denotes the top-KsubK_sub states under this ordering; if fewer than KsubK_sub evaluated states exist, all evaluated states are retained. An empty evaluated pool or retained set contributes zero to the task-level indicator. The primary task-solving endpoint is TaskSolved(Bsearch)=1||∑t∈[∃i∈t:yi=1],TaskSolved(B_search)= 1|T| _t 1 [∃ i _t:\;y_i=1 ], where yi∈0,1y_i∈\0,1\ is the benchmark-specific success indicator for evaluated state i. Because success-aware terminal ranking places successful candidates before failures, task solving at retained size KsubK_sub is equivalent to this pool-level quantity for any Ksub≥1K_sub≥ 1. Task Success denotes this task-level quantity, whereas yiy_i remains the candidate-level success indicator. Benchmark-level comparisons follow the evaluation protocols of the corresponding benchmarks. For GeomMotif, we report the standard Successful rate using the original benchmark success gate. For MotifBench, we use Task Success as a fixed-search-budget endpoint under Bsearch=100B_search=100 evaluated candidates per task. Additional protocol details, auxiliary benchmark metrics, and clustering-based analyses are reported in Appendix. 5 Experiments 5.1 Experimental Questions and Reporting Policy The experiments address two related questions. First, under the standard GeomMotif protocol, can our framework improve benchmark Successful rates while using the original candidate protocol and success definition? Second, under a fixed candidate/evaluator budget, can ELMS discover successful scaffolds more reliably than non-adaptive generation or simpler iterative feedback search? Benchmark splits, candidate interfaces, and baseline-number sources are summarized in Appendix. Success is the primary main-text endpoint because it is the fixed benchmark-level feasibility question. During search, ELMS uses only candidate-local evaluator feedback and online proxy signals; final benchmark success is recomputed under each benchmark protocol after candidate generation. Novelty is reported post hoc: it is evaluated after the final candidate set is formed and is therefore treated as a characterization metric rather than primary online objectives. 5.2 GeomMotif Success and Novelty Comparison Table 2 compares our framework with published GeomMotif baselines using the original GeomMotif Successful definition and candidate protocol (100 candidates per task), while reporting Novelty as a secondary post hoc statistic. Protocol details are provided in Appendix. ELMS reaches 86.41% Successful on single-motif tasks and 84.57% on paired-motif tasks, while also reaching 83.3% Novel on single-motif tasks and 65.7% Novel on paired-motif tasks. These gains substantially exceed the strongest baseline methods on both single- and paired-motif tasks, indicating that evaluator-guided search improves feasibility across GeomMotif constraint settings. We also include a Baseline LLM that directly revises the current sequence from candidate-local evaluator feedback, without the action library, branch preservation, UCB selection, or value back propagation. Table 2: GeomMotif success and novelty comparison. All rows use the standard GeomMotif protocol (100 candidates per task) and benchmark Successful definition. Values are reported as mean ± standard deviation. Post hoc Unique and additional diversity statistics are reported in Appendix. Model / system Successful (%) Novel (%) Single Paired Single Paired Genie2 60.1±1.060.1± 1.0 32.9±0.432.9± 0.4 60.1±1.060.1± 1.0 26.6±0.526.6± 0.5 La-Proteina 67.1±0.767.1± 0.7 62.7±0.562.7± 0.5 67.1±0.667.1± 0.6 35.2±0.935.2± 0.9 RFdiffusion 65.1±0.465.1± 0.4 43.7±1.043.7± 1.0 65.1±0.465.1± 0.4 25.0±0.525.0± 0.5 Protpardelle-1C 56.2±0.756.2± 0.7 44.6±0.544.6± 0.5 56.2±0.756.2± 0.7 25.2±0.425.2± 0.4 FrameFlow 30.6±0.430.6± 0.4 25.1±0.425.1± 0.4 30.6±0.430.6± 0.4 19.7±0.319.7± 0.3 RFdiffusion2 24.9±0.624.9± 0.6 13.5±0.513.5± 0.5 24.9±0.624.9± 0.6 11.4±0.311.4± 0.3 ESM3 (seq) 17.4±0.317.4± 0.3 6.5±0.56.5± 0.5 11.3±0.511.3± 0.5 0.1±0.00.1± 0.0 DPLM-3B 19.3±0.719.3± 0.7 11.0±0.511.0± 0.5 10.2±0.510.2± 0.5 0.9±0.20.9± 0.2 DPLM-650M 15.9±0.415.9± 0.4 6.5±0.56.5± 0.5 8.7±0.28.7± 0.2 0.4±0.10.4± 0.1 ESM3 (seq & struct) 2.0±0.12.0± 0.1 0.7±0.20.7± 0.2 2.0±0.12.0± 0.1 0.0±0.00.0± 0.0 Baseline LLM (gpt-5-nano) 13.1±0.113.1± 0.1 7.2±0.27.2± 0.2 3.7±0.03.7± 0.0 0.3±0.00.3± 0.0 ELMS (gpt-5-nano) 86.4±0.886.4± 0.8 84.6±0.684.6± 0.6 83.3±0.583.3± 0.5 65.7±0.665.7± 0.6 5.3 MotifBench De Novo Design Comparison MotifBench contains 30 functional motif-scaffolding tasks. We compare methods under a matched 100-candidate search budget per task. Candidates are ranked deterministically for benchmark reporting; because this ranking uses the benchmark success indicator, Task Success is interpreted as whether the 100-candidate pool contains at least one successful candidate. Additional protocol details are provided in Appendix. Table 3 reports the fixed-search-budget MotifBench comparison. Across three runs, ELMS solves 26.7 of 30 tasks on average and achieves an 88.89% Success Rate, exceeding the strongest baseline by 10.7 solved tasks and 35.56 percentage points (16.0 tasks, 53.33%). This large gain shows that evaluator-in-the-loop LLM-guided search is effective at using a fixed search budget to discover successful scaffolds. Table 3: MotifBench fixed-search-budget comparison. Results are averaged over three independent runs; each run evaluates 100 candidates per task. Solutions denotes the mean number of tasks for which at least one evaluated candidate satisfies the MotifBench success criterion. Method Solutions Success Rate(%) RFdiffusion3 15.7 52.22±4.8452.22± 4.84 La-Proteina 16.0 53.33±1.9253.33± 1.92 Protpardelle-1c (c91-large) 14.7 48.89±1.1148.89± 1.11 Genie3 9.0 30.00±1.9230.00± 1.92 Baseline LLM(GPT-5-nano) 8.3 27.78 ± 2.22 ELMS(GPT-5-nano) 26.7 88.89±1.1188.89± 1.11 5.4 Matched-Budget Search Ablation This ablation asks whether the observed GeomMotif Success improvement comes from ELMS rather than from changing the underlying language model. All variants are evaluated on GeomMotif and use the same LLM model (gpt-5-nano), the same blinded motif inputs, the same legal action space, the same structural evaluator, the same Success gate, and the same total candidate/evaluator budget Bsearch=100B_search=100. We instantiate four matched-budget search variants. Random Operator Search samples randomly from the shared legal action library without using evaluator feedback for routing. Generate-then-filter independently generates candidates without feeding evaluator results back into action selection. Greedy Operator Search uses candidate-local evaluator feedback to select operators, but follows only the current trajectory, with no tree selection or value backpropagation. ELMS uses the same feedback and action library while preserving multiple branches, applying UCB selection, and back propagating evaluator-derived values. Unlike Baseline LLM, Greedy Operator Search is therefore a matched-library search framework, differing from ELMS only in its tree-level search machinery rather than its action-selection interface. In Table 4, “Evaluator feedback” indicates whether later action choices can condition on structural evaluator outputs, and “Tree selection / value backpropagation” describes how evaluated feedback is retained and reused. Table 4: Matched-budget GeomMotif search ablation with identical LLM, inputs, evaluator, success gate, and action space, using Bsearch=100B_search=100 evaluated candidates per task. All rows report Task Success: the percentage of tasks with at least one successful candidate. Variant Evaluator feedback Action routing Tree selection / value backpropagation Single Success (%) Paired Success (%) Random Operator Search No Random legal routing No 0.0 0.0 Generate-then-filter No Independent direct generation No 6.1 10.4 Greedy Operator Search Yes Evidence-guided greedy routing No; single trajectory 48.0 53.7 ELMS Yes Evidence-guided action selection UCB selection + value backpropagation 88.6 86.4 5.5 Search-Budget Scaling We next examine how the two feedback-guided search methods use additional inference compute. We fix Ksub=1K_sub=1 and vary the internal search budget over Bsearch∈5,10,20,50,100B_search∈\5,10,20,50,100\. Task Success@1 is the fraction of 57 GeomMotif tasks for which the returned design satisfies the Success gate. Greedy Operator Search and ELMS share the same blinded inputs, evaluator, and legal action space, differing only in whether evaluated alternatives are retained and exploited through tree statistics and value backpropagation. 55101020205050100100404050506060707080809090Search budget BsearchB_searchTask Success@1 (%)Greedy Operator SearchELMS Figure 2: GeomMotif Task Success@1 as a function of internal search budget with Ksub=1K_sub=1. Greedy feedback improves at small budgets but largely plateaus, whereas ELMS continues to benefit from additional search. Figure 2 shows different scaling behavior for the two controllers. Greedy Operator Search rises from 47.37% at B=5B=5 to 59.65% at B=20B=20, but reaches only 63.16% at B=100B=100. ELMS rises from 56.14% to 87.72% over the same budget range. In particular, from B=20B=20 to B=100B=100, greedy search improves by only 3.51 percentage points, while MCTS improves by 24.56 points. The widening gap indicates that the benefit of tree search is not simply due to receiving evaluator feedback: preserving evaluated alternatives allows additional inference compute to continue producing useful exploration rather than repeatedly refining a single trajectory. 6 Conclusion We introduced ELMS for motif scaffolding, which turns structural evaluation into state-specific guidance for iterative protein design. ELMS retains evaluated scaffolds as persistent search states, uses Critic and Policy Agents to interpret structural evidence and select targeted modification strategies, executes these decisions through motif-locked operators, and uses MCTS to adaptively allocate further design effort across promising and repairable states. Across GeomMotif and MotifBench, the framework consistently outperforms the strongest generative baselines under matched candidate/evaluator budgets. These results highlight ELMS as a promising approach for combining structural evaluation, agentic reasoning, and adaptive search in constrained protein design. Future work may incorporate richer structural and functional evidence, extend the action space beyond sequence-level refinement, and evaluate whether evaluator-guided search can generalize to broader protein-design objectives and experimentally validated design settings. AI Use Statement Generative AI tools used for manuscript preparation were used solely to improve the clarity, grammar, and readability of the manuscript. They were not used for research ideation, methodology or experimental design, implementation, or result interpretation. Representative examples of wording-only edits are provided in Appendix. References Ahern et al. (2026) W. Ahern, J. Yim, D. Tischer, S. Salike, S. M. Woodbury, D. Kim, I. Kalvet, Y. Kipnis, B. Coventry, H. R. Altae-Tran, M. S. Bauer, R. Barzilay, T. S. Jaakkola, R. Krishna, and D. Baker Atom-level enzyme active site scaffolding using RFdiffusion2. Nature Methods 23, p. 96–105. External Links: Document Cited by: §1, §1, §2. Browne et al. (2012) C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P. I. Cowling, P. Rohlfshagen, S. Tavener, D. Perez, S. Samothrakis, and S. Colton A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in Games 4 (1), p. 1–43. External Links: Document Cited by: §2, §4.1. Butcher et al. (2025) J. Butcher, R. Krishna, R. Mitra, R. I. Brent, Y. Li, N. Corley, P. T. Kim, J. Funk, S. Mathis, S. Salike, A. Muraishi, H. Eisenach, T. R. Thompson, J. Chen, Y. Politanska, E. Sehgal, B. Coventry, O. Zhang, B. Qiang, K. Didi, M. Kazman, F. DiMaio, and D. Baker De novo design of all-atom biomolecular interactions with RFdiffusion3. bioRxiv. Note: Preprint External Links: Document Cited by: §1. Correia et al. (2014) B. E. Correia, J. T. Bates, R. J. Loomis, et al. Proof of principle for epitope-focused vaccine design. Nature 507, p. 201–206. External Links: Document Cited by: §1, §2. Dauparas et al. (2022) J. Dauparas, I. Anishchenko, N. Bennett, H. Bai, R. J. Ragotte, L. F. Milles, B. I. M. Wicky, A. Courbet, R. J. de Haas, N. Bethel, et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science 378 (6615), p. 49–56. External Links: Document Cited by: §2. Geffner et al. (2026) T. Geffner, K. Didi, Z. Cao, D. Reidenbach, Z. Zhang, C. Dallago, E. Kucukbenli, K. Kreis, and A. Vahdat La-Proteina: atomistic protein generation via partially latent flow matching. Note: Preprint Cited by: §1, §2. Ghafarollahi and Buehler (2024) A. Ghafarollahi and M. J. Buehler ProtAgents: protein discovery via large language model multi-agent collaborations combining physics and machine learning. Digital Discovery 3 (7), p. 1389–1409. External Links: Document Cited by: §2. Gungordu et al. (2026) O. Gungordu, S. Xiong, and F. Fekri PathWise: planning through world model for automated heuristic design via self-evolving llms. In Proceedings of the Forty-third International Conference on Machine Learning, External Links: Link Cited by: §2. Hayes et al. (2025) T. Hayes, R. Rao, H. Akin, N. J. Sofroniew, D. Oktay, Z. Lin, R. Verkuil, V. Q. Tran, J. Deaton, M. Wiggert, R. Badkundri, I. Shafkat, J. Gong, A. Derry, R. S. Molina, N. Thomas, Y. A. Khan, C. Mishra, C. Kim, L. J. Bartie, M. Nemeth, P. D. Hsu, T. Sercu, S. Candido, and A. Rives Simulating 500 million years of evolution with a language model. Science 387 (6736), p. 850–858. External Links: Document Cited by: §2. Jiang et al. (2008) L. Jiang, E. A. Althoff, F. R. Clemente, L. Doyle, D. Röthlisberger, A. Zanghellini, J. L. Gallaher, J. L. Betker, F. Tanaka, I. Barbas, D. Hilvert, K. N. Houk, B. L. Stoddard, and D. Baker De novo computational design of retro-aldol enzymes. Science 319 (5868), p. 1387–1391. External Links: Document Cited by: §2. Jumper et al. (2021) J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Zidek, A. Potapenko, et al. Highly accurate protein structure prediction with AlphaFold. Nature 596 (7873), p. 583–589. External Links: Document Cited by: §2. Kocsis and Szepesvári (2006) L. Kocsis and C. Szepesvári Bandit based monte-carlo planning. In Machine Learning: ECML 2006, Lecture Notes in Computer Science, Vol. 4212, p. 282–293. External Links: Document Cited by: §2, §4.1. Kuhlman et al. (2003) B. Kuhlman, G. Dantas, G. C. Ireton, G. Varani, B. L. Stoddard, and D. Baker Design of a novel globular protein fold with atomic-level accuracy. Science 302 (5649), p. 1364–1368. External Links: Document Cited by: §2. Lin et al. (2024) Y. Lin, M. Lee, Z. Zhang, and M. AlQuraishi Out of many, one: designing and scaffolding proteins at the scale of the structural universe with Genie 2. arXiv preprint arXiv:2405.15489. Cited by: §1, §2. Lin et al. (2023) Z. Lin, H. Akin, R. Rao, B. Hie, Z. Zhu, W. Lu, N. Smetanin, R. Verkuil, O. Kabeli, Y. Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379 (6637), p. 1123–1130. External Links: Document Cited by: §2, §4.5. Liu et al. (2024) K. Liu, W. Mao, S. Shen, X. Jiao, Z. Sun, H. Chen, and C. Shen Floating anchor diffusion model for multi-motif scaffolding. In Proceedings of the 41st International Conference on Machine Learning, Vol. 235, p. 31691–31708. Cited by: §1, §2. Röthlisberger et al. (2008) D. Röthlisberger, O. Khersonsky, A. M. Wollacott, L. Jiang, J. DeChancie, J. Betker, J. L. Gallaher, E. A. Althoff, A. Zanghellini, O. Dym, S. Albeck, K. N. Houk, D. S. Tawfik, and D. Baker Kemp elimination catalysts by computational enzyme design. Nature 453 (7192), p. 190–195. External Links: Document Cited by: §2. Strashnov et al. (2026) P. Strashnov, A. Shevtsov, V. Meshchaninov, O. Kardymon, and D. P. Vetrov GeomMotif: a benchmark for arbitrary geometric preservation in protein generation. Note: Preprint Cited by: §1, §2, §3.1. Wang et al. (2022) J. Wang, S. Lisanza, D. Juergens, D. Tischer, J. L. Watson, K. M. Castro, R. Ragotte, A. Saragovi, L. F. Milles, M. Baek, et al. Scaffolding protein functional sites using deep learning. Science 377 (6604), p. 387–394. External Links: Document Cited by: §1, §2. Wang et al. (2024) X. Wang, Z. Zheng, F. Ye, D. Xue, S. Huang, and Q. Gu Diffusion language models are versatile protein learners. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, p. 52309–52333. Cited by: §2. Watson et al. (2023) J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe, J. Yim, H. E. Eisenach, W. Ahern, A. J. Borst, R. J. Ragotte, L. F. Milles, et al. De novo design of protein structure and function with RFdiffusion. Nature 620 (7976), p. 1089–1100. External Links: Document Cited by: §1, §1, §2. Xiong et al. (2026a) J. Xiong, I. Gaur, M. Lukarska, H. Nisonoff, L. M. Oltrogge, D. F. Savage, and J. Listgarten Property guidance for protein sequence generative models with ProteinGuide. Nature Biotechnology. External Links: Document Cited by: §2. Xiong et al. (2026b) S. Xiong, O. Gungordu, J. C. Kerce, and F. Fekri Adaptive information control for search-augmented LLM reasoning. arXiv preprint arXiv:2602.01672. External Links: 2602.01672, Link Cited by: §2. Xiong et al. (2026c) S. Xiong, A. Payani, and F. Fekri Enhancing language model reasoning with structured multi-level modeling. Note: Preprint Cited by: §2. Yim et al. (2024) J. Yim, A. Campbell, E. Mathieu, A. Y. K. Foong, M. Gastegger, J. Jiménez-Luna, S. Lewis, V. Garcia Satorras, B. S. Veeling, F. Noé, R. Barzilay, and T. S. Jaakkola Improved motif-scaffolding with SE(3) flow matching. Transactions on Machine Learning Research. Cited by: §1, §2. Zheng et al. (2025) Z. Zheng, B. Zhang, K. Didi, K. K. Yang, J. Yim, J. L. Watson, H. Chen, and B. L. Trippe MotifBench: a standardized protein design benchmark for motif-scaffolding problems. arXiv preprint arXiv:2502.12479. External Links: 2502.12479 Cited by: §1, §2, §3.1.

Instant research discovery