acceptodds
Under review as a conference paper at ICLR 2027

From Evidence to Action: Experience- and Knowledge-Grounded Evolution for Materials Discovery

Abstract

Iterative materials discovery entails a loop of evaluating generated candidates and feeding the results back into search. Yet evaluation alone does not determine which evidence a task should use or which action a changing search state should trigger. We study this evaluation-to-action gap in heterogeneous, budget-constrained materials discovery. We propose Evidence-to-Action (E2A), an evidence-grounded decision framework for iterative materials discovery that converts internal search experience and retrieved external knowledge into task-conditioned workflows and state-conditioned search actions. The Workflow Policy selects an evidence-use pathway once per run, while within the Full-evidence pathway, the Evolution Policy periodically selects an executable search action from the available state and evidence. In experiments across fourteen tasks, E2A improves hit rate and stable-success rate on eleven tasks each under matched GPT-4o run-native evaluation. Task-conditioned routing exceeds every fixed workflow on the historical return table, while a matched three-task controller study shows task-dependent gains from learned state-conditioned search actions. Together, these findings show how evaluated outcomes become useful through task-appropriate evidence pathways and state-conditioned actions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
materials discovery, large language models, closed-loop discovery, reinforcement learning, retrieval-augmented generation, autonomous scientific discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iTDoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Evidence-to-Action (E2A), a framework for iterative materials discovery that integrates historical search experience and retrieved external knowledge to guide subsequent search decisions. The framework uses two learned policies: a Workflow Policy that selects an evidence-use workflow for each task, and an Evolution Policy that dynamically chooses among exploration, refinement, and escape actions based on the search state. The authors evaluate E2A on 14 materials-design tasks and report improvements in hit rate and stable-success rate over independent generation. Additional experiments compare fixed workflows and learned versus rule-based search controllers.

Critical strengths

The paper addresses a meaningful problem in iterative materials discovery: how to translate accumulated search experience and external knowledge into effective search decisions. The hierarchical design separating workflow selection from search-action selection is reasonable.

The evaluation covers 14 materials-design tasks and multiple LLM backbones. The reported improvements in material discovery performance are encouraging, and the authors provide additional workflow and controller comparisons to investigate the contributions of different components.

Critical weaknesses

The proposed framework mainly combines existing techniques, including experience-based feedback, retrieval-augmented generation, offline policy learning, and evolutionary search control. The two-level policy design is reasonable, but the underlying technical challenges and methodological innovations appear limited. It is unclear what fundamentally new algorithmic insight is introduced beyond integrating these existing components.

The paper does not clearly demonstrate whether the learned policies generalize to unseen materials-design tasks. Since the Workflow Policy is trained using logged outcomes from a small set of tasks, it is unclear whether the model learns transferable decision strategies or task-specific workflow preferences. Moreover, the learned Evolution Policy shows only marginal utility improvements over the rule-based controller on two of the three evaluated tasks (Table 10), and the comparisons use only a single random seed. These limitations weaken the evidence supporting the effectiveness of the learned policies.

Questions for authors

How are the training and evaluation tasks separated for the Workflow Policy? Can the authors provide task-disjoint experiments, such as leave-one-task-out evaluation, to demonstrate generalization to unseen materials-design tasks?

What is the main technical novelty compared with existing experience-learning, adaptive retrieval, and evolutionary search methods? Why are learned LLM-based policies necessary instead of simpler rule-based or lightweight classification approaches?

In Table 10, the utility improvements of the learned Evolution Policy are very small for two tasks. Can the authors provide results across multiple random seeds to establish whether these improvements are statistically reliable?

Since E2A introduces additional policy inference and knowledge retrieval, what is the computational overhead compared with the baselines? Would the improvements remain under a matched total computational budget?

4 comments
AuthorsOpenReview ↗

Response to Reviewer iTDo (1/2)

We thank the reviewer for the thoughtful assessment and constructive questions.

1. Methodological Novelty and Learned Decision-Making (W1, Q2)

We agree that E2A builds on established techniques, including experience feedback, retrieval, offline learning, and evolutionary search. The central methodological question we address, however, is not simply how to incorporate these components, but how accumulated evidence should guide decisions that change the subsequent search process.

Experience feedback and retrieval can provide useful information to a generator, but selecting the appropriate evidence for a task and deciding how to respond to changing search conditions are distinct problems. E2A makes these decisions explicit at two timescales: the Workflow Policy selects an evidence-use configuration before each run, while the Evolution Policy uses the current search state and evidence to select executable EXPLORE, REFINE, or ESCAPE interventions during search. The latter affects candidate selection and search targets, rather than merely adding more text to the generation prompt.

The importance of this distinction is also reflected in our controlled comparisons. With learned action selection disabled, adding Dynamic Knowledge improves discovery on two of the three evaluated tasks, showing that retrieved evidence can contribute independently of the learned controller. Conversely, when the full-evidence workflow is fixed, trained Qwen outperforms both the rule-based controller and unfinetuned Qwen on all three tasks, showing an additional contribution from learned action selection.

We also tested whether an LLM-based controller is necessary by comparing it with rule-based and lightweight alternatives under the same full-evidence workflow and 300-round search budget.

PolicyPVSSEStructural
Trained Qwen0.062580.763480.04286
Rule0.048830.701560.03418
Raw Qwen0.030110.673030.02487
Logistic Regression0.028080.534480.06228
GBDT0.060430.577770.06244
Numerical Critic0.042450.693090.03653

Discovery Utility from matched 300-round active-search comparisons with one seed. Logistic regression and GBDT use an 18-feature numerical representation, while Qwen uses structured textual evidence.

Trained Qwen achieves the highest Discovery Utility on Photovoltaic and solid-state electrolytes, while lightweight classifiers perform better on Structural materials. These comparisons demonstrate that the effectiveness of learned search-action selection varies across tasks and policy models.

The contribution of E2A is therefore the evidence-to-action formulation: separating task-level evidence selection from state-level search intervention, making both decisions learnable, and connecting their outputs to executable changes in a closed-loop discovery process. The individual retrieval and evolutionary operators are established techniques; the methodological contribution lies in how their use is decided and evaluated.

AuthorsOpenReview ↗

Response to Reviewer iTDo (2/2)

2. Generalization and Reliability of the Learned Policies (W2, Q1, Q3)

Workflow Policy generalization (Q1). In the original evaluation (Section 4.3), the Workflow Policy learns from historical returns across the same fourteen-task collection used for evaluation. This tests whether task experience can inform workflow selection, but does not establish generalization to entirely unseen tasks.

To address this concern, we conducted a fourteen-fold task-return-held-out evaluation. In each fold, the target task's returns are excluded from teacher fitting, class weighting, and policy training. The policy then selects a workflow based on the task specification, and this choice is evaluated using frozen historical returns rather than a new online search.

The learned router outperforms fixed M3 on seven held-out tasks, matches it on three, and underperforms on four, including a substantial loss on solid-state electrolytes. These results indicate that workflow preferences can transfer beneficially to some tasks without their historical return labels, while the effectiveness of such transfer remains task-dependent.

Evolution Policy reliability (Q3). The reviewer raises an important question about whether the small improvements reported in Table 10 persist beyond the original random seed. We therefore added two matched seeds for each of the three controller tasks, keeping the full-evidence workflow and search budget fixed.

TaskQwen − Rule Discovery Utility (Mean ± SD)
Photovoltaic+0.01162 ± 0.02562
Solid-state Electrolytes+0.14016 ± 0.03815
Structural Materials+0.00648 ± 0.00514

Paired differences over two additional seeds per task, with 300 search rounds per run; SD denotes sample standard deviation.

Five of the six additional paired runs favor the learned controller, with positive mean improvements across all three tasks. The gains are particularly pronounced on solid-state electrolytes, while Photovoltaic and Structural materials also show improvements. These results further support the effectiveness and reproducibility of learned action selection beyond the original single-seed evaluation.

3. Computational Overhead and Budget Matching (Q4)

Computational Cost. The published baselines include conventional crystal generators such as CDVAE, G-SchNet, DiffCSP, and MatterGen, whose computational costs cannot be directly compared using the token-based accounting applicable to LLM-based generation. We therefore focus on the direct comparison with Independent GPT-4o, using the same generation backbone and 300 search rounds per task.

E2A introduces additional inference costs primarily through richer input contexts, with 95.05% of the additional recorded tokens coming from input. Under a common reference-price scenario, the estimated GPT-4o inference expenditure across all 14 tasks is $58.98 for E2A versus $41.06 for Independent GPT-4o. This corresponds to an additional $17.92 in total, or approximately $1.28 per task on average. Meanwhile, the aggregate observed runtime remains comparable (75,784 s for E2A versus 78,418 s for independent generation).

These results show that E2A achieves the reported improvements in materials-discovery performance with a modest additional GPT-4o inference expenditure per task, without increasing observed task-run time.

AuthorsOpenReview ↗

A Gentle Follow-up on Further Feedback

Dear Reviewer,

Thank you very much for your time and valuable feedback.

As the discussion period is approaching its end, with fewer than three days remaining, we would kindly like to ask whether you have any further questions or concerns regarding our responses. We would be more than happy to provide additional clarification if needed.

We sincerely appreciate your time, consideration, and constructive feedback.

Best regards, The Authors

Thanks for the detailed response and the additional experiments.

The comparisons with simpler controllers and the additional matched-seed results are helpful. They address some of my concerns about the evolution policy, although the improvements are still relatively small on some tasks.

I also appreciate the task-return-held-out evaluation. The results are mixed though, as the learned workflow policy performs worse than fixed M3 on four tasks. Also, since the evaluation uses historical returns rather than new search runs, I'm still not fully convinced about its generalization to unseen tasks.

As for the novelty, I understand that the main contribution is the overall framework rather than the individual algorithms. Still, I think the technical novelty is somewhat limited, since most of the components are based on existing techniques.

The computational cost analysis is useful too, but using the same number of search rounds doesn't necessarily mean the computational costs are the same.

All in all, some of my concerns have been addressed, but I still have reservations about the technical novelty and generalization.

Reviewer_sAN4OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript explores the “evaluation-to-action gap” in materials discovery, that is, evaluation results cannot automatically determine which evidence to use or which search action to trigger. The authors propose the E2A framework, which contains two levels of learned decision-making: (1) before search, a Workflow Policy selects and fixes a single evidence workflow (M1–M3); and (2) under the Full-evidence workflow, an Evolution Policy selects a search action (EXPLORE, REFINE, or ESCAPE) from the state and evidence every three iterations. Across 14 tasks, E2A improves the hit rate and success rate on 11 tasks under the same budget.

Critical strengths
  1. The paper reframes materials discovery as a two-level decision problem, namely “which evidence to use” and “how to convert it into action,” resulting in a clear framework.

  2. Compared with independent generation, this framework significantly improves the success rate on most tasks; for example, solid-state electrolytes improve from 0.52%(0.00%) to 89.37%(85.32%).

Critical weaknesses
  1. The paper only states that a materials-search schema is applied, but does not explain whether the selection is rule-based or summarized by an LLM. For example, what counts as near-feasible? What counts as a bottleneck?
  2. Could the authors provide results for “M2 + retrieval − Evolution Policy”?
  3. For the sake of rigor, I believe the following sensitivity analyses should be added: (1) the reward weights; (2) the action dwell length h; (3) the weight of the property signal in retrieval, λp; (4) a comparison between BM25 alone and BM25 with facet reranking; and (5) the length limit of the Prompt Feedback.
  4. Among the three backbones, only GPT-4o has a matched baseline. The other two backbones have no control group, so it is impossible to determine whether the improvements come from the framework or from the backbone itself.
  5. What is the reason for using only Qwen3-8B for the policies?
  6. Every setting in the paper uses a single seed, with no variance reported. The authors should run multiple seeds and report the average, which would make the experiments more convincing; otherwise, randomness adds yet another source of uncertainty to the framework.
  7. Which policy generated the offline data, i.e., what is the behavior policy?
  8. Both the Workflow Policy and the Evolution Policy are trained on very few samples. How can the authors show that the models are not simply memorizing the answers? I suggest adding a leave-one-task-out experiment to enable a fairer comparison.
  9. Although such benchmarks may be lacking in the materials domain, I still recommend applying your framework to similar tasks in other domains.
  10. I am curious how the reward weights were designed.
  11. The paper mentions an 18-feature classical projection. If logistic regression or GBDT were used on State v3 to make the same choices, how would their performance compare with the LLM?
  12. How many of the surrogate-model hits would pass DFT validation?
  13. My major concern is that the baselines used for comparison are not convincing enough, because the comparison with LLEMA uses different evaluation methods and constraints. Could the authors design stronger baselines for comparison?
Questions for authors

As mentioned in the Weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer sAN4 (1/3)

We thank the reviewer for the detailed and constructive comments. We have conducted additional controlled experiments to examine the contributions of different components, the effectiveness of learned policies, and the reliability of the reported results. We also provide further clarification of the training and evaluation procedures below.

1. Search Schema and Component Attribution (Q1–Q2)

Search schema (Q1). The materials-search schema is constructed deterministically from evaluated search trajectories rather than summarized by an LLM. Specifically, under the REFINE action, the highest-scoring unsuccessful candidate is selected as the near-feasible parent. The main bottleneck is identified from recent constraint violations: the most frequently violated constraint is designated as the main bottleneck when its share reaches 30%. These quantities are computed before policy inference, and the selected action is subsequently applied by the search executor.

Component attribution (Q2). We conducted the requested comparison by adding Dynamic Knowledge to Structured Experience Feedback while keeping the learned Evolution Policy disabled. Both conditions use the same non-learning BaselineController, which preserves the original search execution without policy-selected action overrides. The generator, initial candidate pool, evaluator, and 300-round search budget are matched within each task.

TaskStructured EFStructured EF + Dynamic Knowledge
Photovoltaic (PV)0.034800.02808
Solid-state Electrolytes (SSE)0.107550.76932
Structural Materials0.034860.04344

Discovery Utility under the same non-learning BaselineController, with matched task-specific initial pools, generator, seed, and 300 search rounds. The second condition adds description-first Dynamic Knowledge without enabling learned Evolution.

Dynamic Knowledge improves Discovery Utility on SSE and Structural under the non-learning controller. This comparison allows us to examine the contribution of external knowledge without simultaneously introducing learned action selection.

Conversely, the learned-versus-rule controller experiments hold the full-evidence workflow fixed and examine the contribution of action selection separately. Together, these comparisons provide complementary evidence for knowledge retrieval and learned search decisions beyond the original M2-to-M3 comparison.

2. Sensitivity Analysis (Q3)

We conducted additional analyses covering the five factors suggested by the reviewer: reward weights, action dwell length, retrieval property weighting, BM25 with facet reranking, and feedback length.

FactorTested settingsMain result
Reward weightsSubmitted weights and three alternative profiles87.1%–90.2% agreement with reference offline ridge decisions
Action dwell length1 / 3 / 6 roundsPV Utility: 0.03992 / 0.02311 / 0.03773
Retrieval property weightλp = 0 / 0.1 / 0.2 / 0.4 / 0.6Top-3 property fit: 0.7145 → 0.8286
BM25 + facet rerankingFixed retrieved candidate poolsMean Top-3 property fit: +0.0168 over text-score-only ranking
Feedback length800 / 1600 / 2400 charactersPV Utility: 0.04122 / 0.03294 / 0.02330

The reward-weight and retrieval-weight analyses are offline diagnostics, while action dwell and feedback length are evaluated through active search on PV. The dwell experiments use the same frozen policy without retraining for each setting.

For retrieval, we additionally evaluated BM25 and description-first retrieval in active discovery. The fixed-candidate-pool analysis isolates the effect of facet reranking on property fit, while the active comparison examines the complete retrieval strategies during search.

AuthorsOpenReview ↗

Response to Reviewer sAN4 (2/3)

3. Generation Backbones and Policy Model (Q4–Q5)

Generation backbones (Q4). Our primary experiments are designed to evaluate whether evidence organization and search decisions improve discovery under controlled generation conditions. Accordingly, the framework comparisons use the same generator, task constraints, evaluator, and search budget, allowing us to distinguish the contribution of the framework from differences in backbone capability.

We also conducted a matched comparison between GPT-4o and GPT-5-mini across fourteen materials-design tasks while keeping the framework configuration fixed. This comparison examines how the generator influences discovery quality and exploration behavior, complementing the fixed-generator framework evaluations.

Policy model (Q5). We selected Qwen3-8B because it supports parameter-efficient local adaptation, structured textual evidence as input, and constrained discrete action outputs. This makes it suitable for implementing both decision policies within the proposed framework.

We further evaluated this choice through comparisons with unfinetuned Qwen and lightweight decision models, including logistic regression and GBDT. The corresponding active discovery results are presented under Q11.

4. Reliability, Data Collection and Generalization (Q6–Q8)

Multi-seed reliability (Q6). To examine the reliability of learned action selection, we extended the original three-task controller comparison in Section 4.4 with two additional matched seeds per task. This preserves the original task selection and evaluation protocol while allowing us to investigate variability across independent runs.

TaskQwen − Rule Discovery Utility (Mean ± SD)
Photovoltaic+0.01162 ± 0.02562
Solid-state Electrolytes+0.14016 ± 0.03815
Structural Materials+0.00648 ± 0.00514

Paired Discovery Utility differences over the two additional seeds per task, using 300 search rounds. SD denotes sample standard deviation.

Five of the six paired runs favor the learned controller. The improvement is most pronounced on SSE, with smaller positive mean differences on PV and Structural. These results extend the original single-seed comparison and provide additional evidence for the effectiveness of learned action selection.

Behavior policy (Q7). The Workflow Policy is trained using logged outcomes from fixed-workflow runs. For the Evolution Policy, offline trajectories were collected using a randomized behavior policy that assigns EXPLORE, REFINE, and ESCAPE in balanced three-round action blocks, with the corresponding action-selection probabilities recorded.

The final Evolution adapter uses 60 retained positive-advantage records for training and 135 separate validation inputs. Policy observations contain only information available before each decision, while subsequent evaluation outcomes are used to construct the training targets. This separates the evidence available at decision time from the feedback used for policy learning.

Workflow generalization (Q8). The original evaluation in Section 4.3 examines whether historical task returns can guide workflow selection across heterogeneous materials-design tasks, where learned routing outperforms the strongest fixed workflow.

To further investigate generalization, we conducted a fourteen-fold task-return-held-out evaluation. Each fold excludes the target task's returns from teacher fitting, class weighting, and policy training. The learned policy selects a workflow based on the task specification, and the selected workflow is evaluated using frozen historical returns rather than a new online search.

Across the fourteen held-out tasks, the learned router outperforms fixed M3 on seven tasks, matches it on three, and underperforms on four. The outcomes vary across tasks, including a substantial loss on solid-state electrolytes. These results provide a more detailed view of workflow selection when target-task return information is unavailable, complementing the historical routing evaluation in the original paper.

AuthorsOpenReview ↗

Response to Reviewer sAN4 (3/3)

5. Further Methodological Questions (Q9–Q12)

Cross-domain applicability (Q9). Evaluating E2A beyond materials discovery would be valuable for understanding whether the two-level decision framework transfers to other iterative scientific search problems.

Our current evaluation covers fourteen heterogeneous materials-design tasks involving different property objectives and search constraints. These experiments allow us to study the framework across diverse problems within materials discovery.

Reward design (Q10). The reward is designed to reflect several complementary objectives in iterative materials search. It assigns weights to margin improvement (0.30), new unique stable successes (0.25), constraint progress (0.15), archive novelty (0.10), and diversity (0.10), together with penalties for duplicate, invalid, and repeatedly evaluated candidates.

This design combines successful discoveries with intermediate progress toward feasibility and continued exploration. The intermediate signals are particularly useful when stable successes are sparse, as they provide feedback on whether a search action is moving toward promising regions. The reward-weight sensitivity results are provided under Q3.

Lightweight policy comparisons (Q11). To assess whether the learned LLM policy offers advantages over simpler decision models, we compared trained Qwen with logistic regression, GBDT, unfinetuned Qwen, a numerical critic, and the rule-based controller under the same full-evidence workflow and 300-round search budget.

PolicyPVSSEStructural
Trained Qwen0.062580.763480.04286
Rule0.048830.701560.03418
Raw Qwen0.030110.673030.02487
Logistic Regression0.028080.534480.06228
GBDT0.060430.577770.06244
Numerical Critic0.042450.693090.03653

Discovery Utility from active discovery experiments with one matched seed. Raw Qwen denotes the unfinetuned model.

Trained Qwen outperforms both Rule and unfinetuned Qwen on all three evaluated tasks. Logistic regression and GBDT use the same 60 retained training records through an 18-feature numerical representation, whereas Qwen uses structured textual evidence. The lightweight models achieve higher Utility on Structural materials.

These comparisons evaluate the consequences of the policy decisions during actual materials search, rather than relying solely on offline action-prediction accuracy. The results support the contribution of learned action selection under the shared evidence and search mechanisms.

DFT validation (Q12). Our study focuses on evaluating search decisions under a consistent computational screening protocol, where candidate properties are assessed using the same surrogate models and task constraints across methods.

DFT validation would provide an additional level of physical verification for the discovered candidates. Since the present study evaluates the search framework rather than reporting experimentally or DFT-confirmed materials, we do not report a DFT pass fraction. We agree that such validation would be valuable for subsequent materials-focused investigations.

6. Baseline Fairness (Q13)

We agree that results obtained using different property evaluators or task constraints should not be treated as directly matched comparisons.

Our main conclusions are supported by two complementary experimental settings.

First, the fourteen-task comparison between independent GPT-4o generation and E2A uses the same generator, task constraints, evaluation protocol, seed, and 300-round search budget. This evaluates the contribution of the complete E2A framework relative to independent generation.

Second, the learned-versus-rule controller experiments hold the full-evidence workflow and search mechanisms fixed, examining the additional contribution of learned action selection. The new six-policy comparisons further strengthen this evaluation.

The published LLEMA results provide broader context about existing materials-discovery methods, while our matched experiments directly evaluate E2A's contributions under consistent conditions. We have also compared different generation backbones and included specialist generators such as MatterGen and DiffCSP++ as additional references. These comparisons examine different aspects of generation capability and are kept distinct from the controlled framework evaluations.

Reviewer_1JdrOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the E2A framework, which translates past experience and external knowledge into concrete search decisions by selecting evidence sources based on the task and choosing exploration or optimization actions according to the current search state, thereby improving the success rate of materials discovery under a fixed iteration budget.

Critical strengths
  1. The paper is motivated by the observation that evaluation results do not automatically translate into good search decisions. This motivation is reasonable, and the methodological design clearly addresses it.
  2. The full system shows clear gains in paired experiments. The evaluation covers a broad range of tasks, comparing HR/SR (%) across fourteen materials-design tasks.
Critical weaknesses
  1. The models used have limited capabilities. Models such as GPT-4o and GPT-5-mini lag considerably behind flagship models in AI4S domains such as materials discovery. GPT-4o, in particular, dates back to 2024.
  2. Key modules are bundled together, and the ablation studies are insufficient. The transition from M2 to M3 introduces both dynamic knowledge and an action policy, making it impossible to isolate their individual contributions. There are also insufficient comparisons between structured experience and length-matched raw histories, dynamic and static retrieval, and description-first and standard retrieval.
  3. The necessity and generalizability of the learned policy remain unproven. The Evolution adapter is ultimately trained on only 60 inputs, and comparisons against baselines such as simple classifiers and LLMs without fine-tuning are missing.
Questions for authors

See the Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1Jdr

We thank the reviewer for the constructive comments on the generation backbones, component attribution, and learned policies. We have conducted additional controlled experiments to address these concerns and clarify the contributions of E2A.

1. Generation Backbone

Our study focuses on whether evidence organization and search decisions improve materials discovery, rather than benchmarking the latest LLMs. Given the substantial computational cost of evaluating additional models across fourteen tasks with 300 search rounds each, we prioritized controlled comparisons using GPT-4o and GPT-5-mini.

2. Component Attribution and Ablations

As the reviewer correctly points out, the M2-to-M3 comparison introduces both Dynamic Knowledge and the Evolution Policy. However, the submitted paper already includes a separate comparison of learned and rule-based controllers in Section 4.4 and Appendix Table 10, where the full-evidence workflow and search settings are held fixed.

To further distinguish the contributions of experience organization and external knowledge, we conducted additional controlled experiments on Photovoltaic (PV), solid-state electrolytes (SSE), and Structural materials. The following table reports task-level differences in Discovery Utility.

Comparison (Δ Discovery Utility)PVSSEStructural
Structured EF − Raw History+0.02162+0.02054+0.01438
Static Snippets − No Retrieval−0.01679−0.10531−0.00488
Description-first − Dynamic BM25−0.01421+0.59814+0.02087
Dynamic BM25 − Fixed Initial Cards+0.02561−0.05260+0.00283

Task-level differences in Discovery Utility under matched 300-round, single-seed comparisons. The feedback comparison excludes retrieval and learned Evolution; retrieval comparisons use the non-learning BaselineController. Static Snippets are fixed BM25 document excerpts, whereas Fixed Initial Cards preserve the same initial knowledge-card text as Dynamic BM25. The last two comparisons reuse the same Dynamic BM25 runs. Each row represents a separate controlled comparison, not an additive decomposition of component contributions.

Under the same feedback-token ceiling, Structured Experience Feedback improves Discovery Utility over raw history on all three tasks, while using 2.96% fewer estimated prompt tokens. With retrieval and learned Evolution disabled, this comparison supports the independent contribution of organizing evaluated search experience.

The retrieval experiments further distinguish knowledge selection from updating. Description-first retrieval improves over dynamic BM25 on SSE and Structural, while BM25 performs better on PV. The fixed-initial-card control separately examines the effect of refreshing knowledge during search.

3. Learned Policy

The Evolution Policy addresses a relatively simple and well-defined decision task: selecting among three predefined search actions (EXPLORE, REFINE, and ESCAPE), rather than generating materials directly. With only 60 selected training records, the trained Qwen policy already achieves strong performance in active materials discovery.

To evaluate the learned policy, we compared trained Qwen with unfinetuned Qwen, logistic regression, GBDT, a numerical critic, and the rule-based controller under the same full-evidence workflow and search settings.

PolicyPVSSEStructural
Trained Qwen0.062580.763480.04286
Rule0.048830.701560.03418
Raw Qwen0.030110.673030.02487
Logistic Regression0.028080.534480.06228
GBDT0.060430.577770.06244
Numerical Critic0.042450.693090.03653

Discovery Utility from 300-round active discovery experiments with one matched seed. Raw Qwen denotes the unfinetuned model.

Trained Qwen outperforms both the rule-based controller and unfinetuned Qwen on all three tasks, achieving the highest Discovery Utility on Photovoltaic and solid-state electrolytes. GBDT performs best on Structural materials.

These active-search results show that even with only 60 selected training records, learned action selection provides measurable benefits over fixed rules and the unfinetuned LLM, while remaining competitive with lightweight classifiers across the evaluated tasks.

AuthorsOpenReview ↗

A Gentle Follow-up on Further Feedback

Dear Reviewer,

Thank you very much for your time and valuable feedback.

As the discussion period is approaching its end, with fewer than three days remaining, we would kindly like to ask whether you have any further questions or concerns regarding our responses. We would be more than happy to provide additional clarification if needed.

We sincerely appreciate your time, consideration, and constructive feedback.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions

We have revised the manuscript to clarify the implementation of E2A and provide additional evidence for its experience and knowledge components, task-conditioned workflow selection, and state-conditioned action selection. The main changes are summarized below, with references to the corresponding sections, pages, and appendix tables in the revised manuscript.

1. Clarification of the Evolution Policy and Training Protocol

Locations: Section 2 (p. 3); Section 3.4 (p. 5, lines 247–250); Appendix D.2 (pp. 16–17, Equations 10–11 and Table 7).

We have expanded the explanation of the reward design, clarifying how discovery success, progress toward task constraints, novelty, diversity, and penalties for unproductive search contribute to the training signal.

We have also corrected and clarified the final Evolution Policy training protocol. In particular, the action-execution horizon is three iterations (h=3h=3), whereas the critic and advantage estimates use a nine-iteration return window (H=9H=9). The revised description distinguishes the final training cohort from earlier experiments and specifies the positive-advantage filtering used to select 60 supervised training records. We also clarify the effective supervised QLoRA objective, including the cancellation of per-sample importance weights under the deployed microbatch configuration.

2. Additional Component Attribution Experiments

Locations: Section 4.2 (pp. 6–7, lines 323–327); Appendix G.4 (p. 21, Tables 11–12).

We have added controlled comparisons to examine the contributions of Structured Experience Feedback (EF) and Dynamic Knowledge (DK).

The new experiments compare Structured EF against raw search history under a controlled feedback-token budget, and evaluate alternative retrieval configurations. We also isolate the contribution of Dynamic Knowledge under a fixed, non-learning controller. Structured EF improves Discovery Utility over raw history on all three evaluated tasks, while the effects of retrieval vary across tasks, with a substantial improvement on solid-state electrolytes.

3. Task-Return-Held-Out Workflow Evaluation

Locations: Section 4.3 (p. 7, lines 362–367); Appendix G.5 (p. 22, Table 13).

We have added a fourteen-fold task-return-held-out evaluation of the Workflow Policy. In each fold, the target task's historical workflow returns are excluded from policy training, and the selected workflow is evaluated using frozen historical returns.

Relative to fixed M3, the learned router performs better on seven tasks, matches it on three, and performs worse on four. The results provide a more detailed characterization of task-conditioned workflow selection beyond the original historical routing comparison.

4. Additional Evolution Policy Comparisons and Matched-Seed Evaluation

Locations: Section 4.4 (p. 9, lines 432–437); Appendix G.6 (p. 23, Tables 14–15).

We have expanded the Evolution Policy evaluation to include six action-selection methods: trained Qwen, a rule-based controller, unfinetuned Qwen, logistic regression, GBDT, and a numerical critic.

Under the same Full-evidence workflow, trained Qwen outperforms the rule-based and unfinetuned Qwen baselines on all three evaluated tasks, while lightweight numerical controllers perform better on structural aerospace. We have also added two matched seeds per task for the Learned-versus-Rule comparison, with positive Discovery Utility differences in five of the six additional paired runs.

The appendix specifies the evaluation protocols, random seeds, and treatment of initial-pool candidates for the different experimental panels.

5. Consistency and Presentation Improvements

Locations: Introduction (p. 2); Conclusion (p. 9, lines 441–449); Appendices G.4 (p. 21) and G.6 (p. 23).

We have removed a repeated description of the Workflow Policy in the Introduction and clarified the execution sequence in the Conclusion: workflow selection occurs before a search run, while evidence formation and state-conditioned action selection take place during search.

We have also clarified the feedback-token control used in the EF comparison and the evaluation conventions of the additional controller experiments.

The original main performance tables and action-trajectory figure are retained. Detailed results of the additional experiments are provided in Appendices G.4–G.6, with corresponding summaries in the main text.

These revisions clarify the implemented methodology and strengthen the empirical characterization of E2A's evidence-use and search-decision mechanisms.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.