acceptodds
Under review as a conference paper at ICLR 2027

Learning from Failure: A Physics-Driven Self-Improving Agentic Framework for Liquid Electrolyte Design

Abstract

Liquid electrolytes are foundational to advanced batteries, and their formulation critically affects battery performance, safety, and lifetime. However, designing viable electrolytes remains a longstanding industrial challenge, as it requires navigating an interdependent space of discrete components and continuous concentrations under complex, often conflicting electrochemical constraints. In this work, we introduce SAGE, a Self-improving AGentic framework for Electrolyte discovery that iteratively refines its design policies through physics-driven feedback without human supervision. Central to SAGE is a principled paradigm of self-evolving autonomy, in which the agent iteratively improves through molecular dynamics evaluations and distills case-specific failure experience into actionable guidance for subsequent formulation design. Moreover, SAGE establishes an auditable discovery trajectory that documents constraint violations, retrieved evidence, hypothesis revisions, and resulting policy updates as structured knowledge, providing a reusable blueprint for tackling new scientific problems. Experiments across diverse real-world electrolyte discovery scenarios show the effectiveness of SAGE and the importance of physics-guided self-improvement for designing feasible, novel formulations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Electrolyte Formulation Design, Agentic AI

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_vMu5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors present an agentic framework, SAGE, for designing electrolytes that meet prescribed requirements. They use simulations in the form of molecular dynamics (MD) and density functional theory (DFT) to evaluate performance. New candidates are proposed based on the performance history, using a system of five subagents with different roles. The agents are all pretrained, but have access to different tools and are controlled by different prompts.

Critical strengths

S1: Strong performance

The success rate widely surpasses that of other agentic frameworks, both general-purpose and materials science-specific ones. An ablation study shows that each proposed novelty significantly improves performance.

S2: Clear presentation

The motivation is clear, and each component is presented in a logical fashion.

Critical weaknesses

W1: Overstates the applicability and generalizability

You consistently portray SAGE as being "physics-based" and do not sufficiently state its limitations. Your "physical experiments" are just simulations and would be more accurately described as such; swapping "physics-based" for "simulation-based" would be a good first step. MD and DFT are computationally costly simulations that are certainly informative of electrolyte performance, but at the same time can only probe limited time and spatial scales. Important aspects such as slow reactions, SEI growth, aging, and other degradation phenomena are beyond reach and would require actual physical experiments to test.

W2: Specific insights missing, reproducibility questionable

For something to be a valuable scientific contribution, others must be able to build on top of it.

The proposed agentic framework is highly complex, with many moving parts and intricacies, and it's not clear how it differs from standard harnesses like Claude Code or Codex. Describing all of it in sufficient detail for someone to reproduce would require an excruciatingly detailed paper. As far as I can tell, the only practical way for someone to build on the full framework is if you would open-source the code. Barring that, how is anyone supposed to trust and independently verify that your superior performance actually holds up under scrutiny? (Table 7 provides an ablation, but I get suspicious when all components provide 25-50% improvements in success rate. Similarly, the improvements in Table 2 seem too good to be true. )

An alternative route would be to isolate and pin down specific mechanisms that are transferable to other settings. Frankly, without proper evidence to back it up, describing the role of each subagent is just storytelling. Take, for instance, the claim that "The reflector routes each actionable failure to the part of the system best suited to address it" (l. 310). This sounds like it would be possible to test, but it currently isn't.

W3: Missing comparisons to non-agentic approaches

In the experiments, baselines only include other agentic frameworks. I think it would be very important to include non-agentic baselines, e.g., Bayesian optimization and even classical Design of Experiments (without feedback). These are probably not applicable in all settings, but please include comparisons where they are applicable. Curiously, in the appendix, the composer is described as using a task-specific objective vector to rank candidates that combines with a genetic algorithm to generate new candidates, so it does seem like you've given some thought to using more classical methods.

Minor: You talk about "gates" long before they are defined.

Questions for authors

Q1: What is specific to electrolytes in their "self-improving agentic search problem"? Why isn't it just an arbitrary black-box optimizer?

Q2: How does SAGE balance exploration vs. exploitation? Why is an agentic approach presumed to be more efficient than established methods such as reinforcement learning, genetic algorithms, or Bayesian optimization? From the looks of it, the setting seems like one where reinforcement learning with verifiable rewards (RLVR) would be applicable, is there a reason for not taking such an approach?

Q3: Why not model all components with nonnegative concentrations, thereby side-stepping the issue of discrete variables?

Q4: How do you deal with the fact that many constraints are not known a priori? See, e.g., Hickman et al. "Anubis: Bayesian optimization with unknown feasibility constraints for scientific experimentation." Digital Discovery 4.8 (2025): 2104-2122.

5 comments

We sincerely thank the reviewer for the careful assessment and constructive suggestions. Below, we address each concern point by point:

W1: Overstates the applicability and generalizability.

We appreciate this important point.

  • We agree that our evaluation is simulation-based rather than based on wet-lab experiments. Nonetheless, MD and DFT provide physically meaningful feedback that can effectively support electrolyte formulation design within the scope of computational screening.
  • We have revised the manuscript title and the relevant text accordingly, with changes marked in red.

W2: Specific insights missing, reproducibility questionable.

This is a very valuable suggestion.

  • We provide additional details of the physical simulation protocols in Appendix A.10. We agree that reproducibility is essential. We have open-sourced our code, including the full agentic workflow, in the supplementary materials, and will release the code for the remaining results upon acceptance.
  • The results in Table 2 and Figure 7 are directly recorded from our experiments. We believe the gains in Table 2 come from SAGE's iterative learning from physical failure feedback through iterative policy updates, whereas the baselines treat evaluations as isolated results. In Figure 7, full SAGE is normalized to 100%, so the relative gaps appear larger than the absolute differences.
  • Unlike general-purpose coding harnesses such as Claude Code or Codex, SAGE is tailored to electrolyte discovery through electrolyte-specific constraints and operators and MD/DFT evaluation of transport, solvation, and redox properties. Its self-improvement strategy translates failed simulation gates into electrolyte-specific diagnoses and actionable formulation strategies, such as changing component identities, ratios, or interactions, and routes them to prompt, skill, and tool-use policies to autonomously improve subsequent search.
  • We add three controlled experiments to isolate the proposed mechanisms: detailed ablations vary whether failure experience is retained and whether it updates prompts, skills, and tool-use (P/S/U); Policy guidance compares fixed, rule-based, and stateless LLM guidance; and failure routing sends the same actionable updates to different policy destinations while keeping all other search components fixed. Together, these experiments test whether the gains come from persistent, actionable updates and correct routing. The results show that policy updates outperform fixed, rule-based, stateless, and read-only alternatives, while incorrect or restricted routing consistently reduces performance.
  • We have included these results in the revised manuscript (Table 4, page 10; Table 12, page 37; Table 13, page 38).

Detailed ablations

AblationP/S/UGeneralFast ChargeLow TempHigh TempHigh Voltage
Fixed Policy0 / 0 / 02930202427
Factual Memory0 / 0 / 03134222630
Instant Reflection0 / 0 / 03436242934
Persistent Reflection0 / 0 / 03943283341
Prompt Only1 / 0 / 02828192726
Skill Only0 / 1 / 03842253138
Tool-Use Only0 / 0 / 13233182331
Prompt + Skill1 / 1 / 04948343650
Prompt + Tool-Use1 / 0 / 13540263240
Skill + Tool-Use0 / 1 / 15158403956
Fixed Failure RulesRule-Based3537253233
Unified ReflectorAdaptive P/S/U4449333546
Read-Only Structured MemoryRead-only P/S; U=04040293542
Shuffled Failure Context1 / 1 / 13744303436
Permuted Operator Weights1 / 1 / 15257394254
Exploration First Selection1 / 1 / 14551364647

Policy guidance machanism

GuidanceGeneralFast ChargeLow TempHigh TempHigh Voltage
Fixed Guidance2832212528
Rule-Based Guidance3536263134
Stateless Reflection3339263037

Failure-routing mechanism

RoutingGeneralFast ChargeLow TempHigh TempHigh Voltage
Random4345353647
Wrong3031192425
Skill Only4246273440
Prompt Only3232212828
Tool-Use Only3537202534

W3: Missing comparisons to non-agentic approaches.

We thank the reviewers for this thoughtful point.

  • We adapt and add eight non-agentic methods: Gaussian Process Bayesian Optimization, static Design of Experiments, NSGA-II, Proximal Policy Optimization (PPO), ExtraTrees active learning, Tree-structured Parzen Estimator (TPE), Thompson sampling, and the Cross-Entropy Method. All methods use the same component pool, admissibility checks, physical evaluation protocol, acceptance thresholds, initialization, and candidate budget, differing only in how subsequent candidates are selected.
  • Classical methods underperform SAGE because they do not explicitly learn and distill physical failure experience into actionable guidance for subsequent search, which is critical in a complex, jointly constrained space.
  • We consider this broader comparison important and have included the results in the revised manuscript, as reported in Table 3 (page 10).
MethodGeneralFast ChargeLow TempHigh TempHigh Voltage
GP Bayesian Optimization191812154
Static DoE138550
NSGA-II222011143
PPO1712794
ExtraTrees Active Learning20179122
TPE1915792
Thompson Sampling1713681
Cross-Entropy Method211910113

Minor.

This is an important point. A gate is a pass/fail criterion for whether a candidate formulation satisfies a predefined property constraint. We have revised the manuscript accordingly on page 3, with changes marked in red.

Q1: Specific design for electrolytes & interpretability.

This is an insightful point.

  • Electrolyte design involves a hybrid discrete-continuous space in which salt, solvent, and additive identities and their concentrations jointly determine coupled properties such as transport, viscosity, solvation, aggregation, and oxidation tendency. Thus, SAGE learns from electrolyte-specific MD/DFT failure signals and updates design policies over chemically meaningful components, interactions, and operators.
  • Rather than treating failure as a scalar reward, SAGE identifies the violated constraint and its formulation context. Failures such as excessive ion aggregation, poor transport, unstable coordination, or unfavorable cluster VIP are then converted into targeted evidence retrieval, hypotheses, and candidate refinement actions.
  • SAGE is not an arbitrary black-box optimizer: it preserves a complete optimization trail from observed failure and supporting evidence to an actionable policy update and subsequent formulation change, while deterministic programs retain numerical authority and fixed physical simulations determine outcomes.
  • Therefore, the resulting artifacts capture meaningful electrolyte design experience rather than black-box objective values, allowing the learned search strategies to be inspected, reused, and adapted across different design regimes in a human-interpretable way.

Q2: Exploration vs. exploitation & efficiency of SAGE & RLVR.

We appreciate these valuable concerns.

  • SAGE balances exploration and exploitation through explicit budget allocation: early iterations emphasize broad coverage, while later iterations exploit validated parents and repair informative near-misses through bounded mutation, substitution, and recombination. Final candidates are selected from a larger pool to balance exploitation, failure repair, evidence support, and compositional diversity.
  • We empirically study exploration-first selection (Table 4, page 10) to assess this trade-off and find that it improves coverage and works particularly well on the high-temperature task, but is less effective on the others, highlighting a task-dependent balance between diversity-driven exploration and targeted exploitation.
  • We do not assume that an agentic approach is inherently efficient. Our added comparisons with classical methods (Table 3, page 10) test this empirically under the same physical budget. SAGE performs better because it improves itself through failure-to-policy learning without any model training and human intervention, whereas the established methods considered mainly rely on numerical rewards or feasibility outcomes.
  • This is a very good point. RLVR is possible in principle, but this setting offers only limited, costly physical simulations, with sparse rewards and delayed MD/DFT feedback under coupled constraints. SAGE instead uses training-free policy refinement, treating individual constraint violations and counterexamples as structured feedback without requiring large-scale policy training or a hand-designed scalar reward. We also included an RL-based approach (PPO) as a baseline in Table 3 (page 10).

Q3: Alternative design of components.

We apprecaite this thoughtful question.

  • Assigning every possible component a nonnegative concentration does not remove the underlying discreteness: component presence changes molecular identity, chemical compatibility, and charge balance, while finite-size MD requires integer molecule counts and cannot faithfully represent arbitrarily small concentrations.
  • Besides, it would instead create a high-dimensional and sparse search space, where many near-zero formulations map to the same realizable molecular composition, reducing both optimization efficiency and interpretability.

Q4: Unknown Priori.

This is a critical concern.

  • SAGE does not require the feasibility boundary to be known a priori: when MD/DFT evaluation reveals an unanticipated failure, SAGE diagnoses the physical cause and converts the evidence into a context-specific policy update that guides subsequent proposals away from similar failures while retaining exploration in unknown regions.
  • In practical electrolyte design, the overall acceptance criteria are predefined [1], but which formulations satisfy them under different conditions is learned through MD/DFT evaluation. Methods such as Anubis are complementary for modeling unknown constraints, whereas SAGE focuses on actionable failure diagnosis and reusable artifacts rather than an explicit probabilistic feasibility model.

[1] Meng, Y. Shirley, Venkat Srinivasan, and Kang Xu. "Designing better electrolytes." Science, 2022.

Reviewer_sREgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAGE enables feedback-driven electrolyte design: It combines language-model reasoning, deterministic composition search, and MD/DFT evaluation to update search policies from failures. Experiments across five computational screening tasks report substantial gains over the evaluated baselines, supported by ablations and within-task policy reuse.

Critical strengths
  1. Clear separation of responsibilities: The framework separates qualitative reasoning, numerical generation, and physical evaluation, supporting traceability and reducing reliance on model-generated judgments.
  2. Encouraging empirical results: The reported improvements across several tasks and model backbones provide evidence of the framework's potential for computational formulation search.
Critical weaknesses
  1. Limited isolation of policy-update benefits: The current experiments leave some uncertainty about how much improvement comes from agent policy updates versus accumulated simulation results and numerical search.
  2. Physical evaluation needs further substantiation: Additional protocol details and robustness analyses would help establish the reproducibility and reliability of the reported pass rates.
  3. Broader claims need clearer qualification: The evidence more directly supports within-task reuse and novelty relative to an indexed corpus; broader transferability and chemical novelty remain less established.
Questions for authors
  1. Contribution of policy updates: Could the authors clarify whether policy updates provide consistent gains when simulation history, numerical search mechanisms, and evaluation budgets are held fixed?
  2. Robustness of the evaluation: Could the authors provide evidence on the stability of pass labels and method rankings across independent runs and reasonable variations in the physical evaluation protocol?
  3. Scope of the broader claims: Could the authors clarify the extent to which the results establish cross-task transferability or chemically distinct formulations beyond within-task reuse and database-level novelty?
4 comments

We sincerely thank the reviewer for the careful assessment and constructive suggestions. Below, we address each concern point by point:

W1 & Q1: Isolated contribution of policy updates.

We thank the reviewers for the insightful sueggestion.

  • We provide ablations (Figure 7, page 9) showing that removing MD feedback, MD archive, agent skills, or tool-use strategies reduces performance. These results quantitatively demonstrate that both accumulated MD feedback and design policy updates contribute to SAGE's performance.
  • We fully agree that an isolated analysis of policy-update benefits is needed. Thus, we add more detailed ablations that keep the simulation history, archive access, parent selection, numerical search, component pool, and evaluation budget fixed, while varying only how design policies (e.g., prompts, skills, and tool-use as P/S/U) are updated and used. We also study policy guidance mechanism by comparing fixed, rule-based, and stateless LLM guidance, while using failure routing to control where identical failure-derived updates are directed. This tests whether policy-update gains depend on both useful guidance and correct routing to design policy components.
  • Empirically, policy updates consistently outperform fixed policy, textual reflection, read-only memory, and failure-conditioned rules, showing that their advantage comes from converting failures into persistent, correctly routed, and actionable policy updates, rather than merely relying on numerical search, textual reflection, or failure-conditioned guidance.
  • We agree that this is a crucial point and have included these results in the revised manuscript (Table 4, page 10; Table 12, page 37; Table 13, page 38).

Isolated ablations

AblationP/S/UGeneralFast ChargeLow TempHigh TempHigh Voltage
Fixed Policy0 / 0 / 02930202427
Factual Memory0 / 0 / 03134222630
Instant Reflection0 / 0 / 03436242934
Persistent Reflection0 / 0 / 03943283341
Prompt Only1 / 0 / 02828192726
Skill Only0 / 1 / 03842253138
Tool-Use Only0 / 0 / 13233182331
Prompt + Skill1 / 1 / 04948343650
Prompt + Tool-Use1 / 0 / 13540263240
Skill + Tool-Use0 / 1 / 15158403956
Fixed Failure RulesRule-Based3537253233
Unified ReflectorAdaptive P/S/U4449333546
Read-Only Structured MemoryRead-only P/S; U=04040293542
Shuffled Failure Context1 / 1 / 13744303436
Permuted Operator Weights1 / 1 / 15257394254
Exploration First Selection1 / 1 / 14551364647

Policy guidance machanism

GuidanceGeneralFast ChargeLow TempHigh TempHigh Voltage
Fixed Guidance2832212528
Rule-Based Guidance3536263134
Stateless Reflection3339263037

Failure-routed policy updates

RoutingGeneralFast ChargeLow TempHigh TempHigh Voltage
Random4345353647
Wrong3031192425
Skill Only4246273440
Prompt Only3232212828
Tool-Use Only3537202534

W2 & Q2: Additional protocol details and robustness analyses.

This is a critical point.

  • We provide additional details of the evaluation protocol in the revised manuscript (Appendix A.10, pages 43--44).
  • Moreover, we repeat the main experiments ten times with different random seeds to assess SAGE's robustness. These runs capture reasonable variation in physical evaluation arising from MD initialization, finite trajectory sampling, cluster extraction, and bootstrap resampling, and show stable performance across baselines and tasks. The results are reported in Table 16 (pages 37--38) of the revised manuscript.
MethodGeneralFast ChargeLow TempHigh TempHigh Voltage
LLMs
GPT-5.6 Terra42.5 ± 10.122.3 ± 6.616.0 ± 5.329.5 ± 8.99.0 ± 5.7
LLaMat-39.8 ± 7.610.1 ± 4.80.4 ± 0.81.1 ± 1.41.8 ± 1.2
NatureLM 8-7B13.9 ± 6.225.9 ± 8.07.8 ± 3.716.6 ± 7.60.3 ± 0.8
DeepSeek V4 Flash16.3 ± 7.329.4 ± 9.47.1 ± 8.314.5 ± 8.90.9 ± 1.3
Qwen3.8-27B13.1 ± 10.47.4 ± 3.51.1 ± 3.60.5 ± 1.21.6 ± 2.2
General-Purpose Agents
GPT-5.6 Terra46.0 ± 9.214.3 ± 7.514.4 ± 5.316.4 ± 9.65.5 ± 3.2
LLaMat-311.9 ± 8.96.1 ± 2.61.1 ± 2.00.4 ± 1.22.1 ± 2.3
NatureLM 8-7B28.3 ± 6.123.6 ± 8.60.5 ± 1.66.5 ± 6.10.6 ± 1.1
DeepSeek V4 Flash27.3 ± 7.933.1 ± 8.917.1 ± 4.815.3 ± 9.13.4 ± 3.5
Qwen3.8-27B18.9 ± 6.025.9 ± 9.00.5 ± 1.62.1 ± 2.40.3 ± 0.8
Adapted Agents
ChemCrow29.4 ± 10.024.9 ± 6.85.4 ± 7.021.5 ± 5.41.0 ± 1.1
ChemGraph34.9 ± 9.210.9 ± 12.07.1 ± 6.121.6 ± 6.25.4 ± 3.2
MatSciAgent21.6 ± 9.415.0 ± 6.93.4 ± 6.914.1 ± 4.61.4 ± 3.9
SAGE77.1 ± 4.278.9 ± 6.163.5 ± 5.159.3 ± 3.385.4 ± 6.7

W3 & Q3: Cross-task transferability & chemical novelty.

These are very valuable points.

  • We further evaluate eight cross-task transfer directions by initializing each target task with artifacts learned from a different control task. Compared with cold-start experiments, transfer improves success rates by 5--14 percentage points, suggesting that SAGE's detailed failure experience can effectively generalize across tasks.
  • We conduct additional experiments to assess chemical novelty using nearest-neighbor distances to a broader formulation reference corpus (e.g., Electrolytomics). This analysis considers ingredient identity, composition ratios, and descriptor-based chemical profiles, going beyond database-match novelty. SAGE achieves the largest distance on all three metrics, supporting its exploration of chemically distinct formulation neighborhoods rather than simple retrieval or recombination of known chemical components.
  • We have included the cross-task transferability and chemical novelty results in the revised manuscript, as reported in Table 15 and Table 14, respectively (pages 37--38).

Cross-task transferability

ControlTransferIteration 2 ControlIteration 2 TransferGain (pp, ↑)
GeneralFast Charge5467+13
Fast ChargeGeneral4961+12
GeneralLow Temp.3044+14
Low Temp.General4958+9
Fast ChargeLow Temp.3037+7
Low Temp.Fast Charge5464+10
Low Temp.High Temp.3745+8
High VoltageHigh Temp.3742+5

Chemical novelty

MethodIngredient Jaccard ↑Composition Total Variation ↑Chemical Profile ↑
DeepSeek V4 Flash0.36270.22460.2591
ChemCrow0.32840.19630.3387
GPT-5.6 Terra0.47160.13980.3524
NatureLM 8-7B0.27410.12760.2148
Qwen3.8-27B0.30180.14690.2326
GP Bayesian Optimization0.21630.13170.1935
SAGE0.63070.36680.4926
AuthorsOpenReview ↗

Kind Reminder: Review for Submission 11841

Dear Reviewer sREg,

I hope this email finds you well. I am writing to gently follow up regarding the review for Submission 11841. We understand that you may have a busy schedule, and we simply wanted to check in, as we have not yet received your review.

If possible, we would be very grateful if you could kindly share an update on your progress or submit your review when you have a moment.

Thank you so much for your time, effort, and valuable support.

Best regards,

Authors of Submission 11841

AuthorsOpenReview ↗

Gentle Follow-Up on Review for Submission 11841

Dear Reviewer sREg,

I hope you're well. A gentle follow-up on Submission 11841: the rebuttal period ends in about a day and we haven't yet received your feedback. We understand you're busy and just wanted to check in, so if possible, a brief update or your review would be greatly appreciated. Thank you for your time and support.

Best regards,

Authors of Submission 11841

Reviewer_b1vLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SAGE, a training-free agentic framework for electrolyte formulation discovery. LLM agents retrieve scientific evidence, generate qualitative hypotheses, and translate simulation failures into persistent updates to agent skills, prompts, and search guidance. A deterministic composition engine generates numerical formulations, while fixed molecular dynamics and density functional theory protocols determine whether candidates satisfy task-specific computational constraints. Across five formulation tasks, SAGE reports an aggregate pass rate of 71.8%, compared with 21.0% for the strongest baseline. Ablation studies and same-task warm-start experiments examine the contributions of feedback and accumulated policy artifacts.

Critical strengths

Explicit separation of reasoning, numerical search, and physical evaluation. SAGE constrains LLMs to qualitative hypothesis generation while assigning numerical composition and acceptance decisions to deterministic programs and physical simulations. Its records preserve formulation-specific failures, successful counterexamples, uncertainty, and provenance. This separation makes the search process inspectable and reduces ambiguity about the basis for accepting a candidate.

Substantial reported improvements across tasks and model backbones. The framework improves computational screening success across five electrolyte-design tasks and multiple LLM backbones. Component ablations and same-task reuse experiments provide supporting evidence that feedback and persistent guidance contribute to the performance of the complete system.

Critical weaknesses

The incremental contribution of structured failure learning is insufficiently isolated. Reflection, persistent memory, prompt adaptation, and LLM-guided scientific optimization have established precedents. SAGE additionally employs an archive-based numerical search with parent selection, mutation, and recombination. The baseline comparisons and ablation descriptions do not clearly establish the benefit of the proposed policy updates beyond that search using fixed guidance or simple failure-conditioned rules. This uncertainty limits the demonstrated methodological novelty and the generalizable contribution to AI. The warm-start experiments establish utility within the same task specification, but do not establish transfer across task regimes.

The physical-evaluation protocol is insufficiently specified for independent verification. The central results depend on MD/DFT-derived acceptance decisions, yet important operational details remain unclear, including viscosity estimation, bootstrap block construction and repetition counts, the numerical pass-confidence threshold, and simulation quality-control criteria. These omissions limit reproducibility and confidence in the reported success rates. The absence of wet-lab validation is an acknowledged limitation; the principal concern is whether the computational claims are adequately supported.

Questions for authors

What specific mechanism distinguishes SAGE’s failure learning from existing reflection-and-memory agents and LLM-guided optimization frameworks? Which ablation isolates this mechanism while preserving archive access, parent selection, numerical search operators, and the physical-evaluation budget?

How does SAGE compare with the same archive-based numerical optimizer using fixed guidance or simple failure-conditioned rules under an identical component pool and evaluation budget? This comparison would clarify the incremental value of LLM-generated hypotheses and persistent policy updates.

Please specify the complete physical-evaluation protocol, including simulation and equilibration settings, viscosity estimation, bootstrap configuration, pass-confidence thresholds, quality-control criteria, and the identity and construction of the fixed task-5 reference.

4 comments

We sincerely thank the reviewer for the careful assessment and constructive suggestions. Below, we address each concern point by point:

W1 & Q1: Incremental contribution of structured failure learning.

This is a very insightful suggestion.

  • We conduct multiple additional ablations that keep the archive-based numerical search, parent selection, mutation, recombination, and evaluation budget fixed, while varying only the guidance and policy-update mechanisms. This setup isolates the contribution of failure learning beyond numerical search, reflection, persistent memory, and simple failure-conditioned rules. Empirically, SAGE’s policy updates outperform fixed guidance, read-only memory, text reflection, and failure-aware rules, showing that structured policy updates provide clear incremental value beyond the underlying search.
  • SAGE compiles constraint-specific physical failures into structured, actionable updates to task prompts, agent skills, and tool-use strategies (P, S, and U, respectively), rather than storing them solely as textual reflection or memory. The Read-Only Structured Memory vs. Prompt + Skill ablation preserves archive access, parent selection, numerical search operators, and the physical-evaluation budget, yet structured policy updates improve PASS counts across all five tasks, showing their incremental contribution.
  • Furthermore, we study a routing mechanism that directs failure-derived updates to the appropriate design policy component. Empirically, correct routing improves performance by ensuring that updates are written to the proper policy component.
  • The scientific value of SAGE lies not simply in its refinement of reflection/memory, agentic search, or LLM-guided optimization, but in how it turns physically verified, context-specific failures into actionable updates to design policies that guide subsequent search behavior. This failure-to-policy strategy provides a physically grounded mechanism for self-improvement beyond fixed guidance, textual memory, or simple failure-conditioned rules.
  • We fully agree on the importance of these controlled studies and have added the corresponding results to the revised manuscript. The results are reported in Table 4 (page 10) and Table 13 (page 38).

Isolated ablations

AblationP/S/UGeneralFast ChargeLow TempHigh TempHigh Voltage
Fixed Policy0 / 0 / 02930202427
Factual Memory0 / 0 / 03134222630
Instant Reflection0 / 0 / 03436242934
Persistent Reflection0 / 0 / 03943283341
Prompt Only1 / 0 / 02828192726
Skill Only0 / 1 / 03842253138
Tool-Use Only0 / 0 / 13233182331
Prompt + Skill1 / 1 / 04948343650
Prompt + Tool-Use1 / 0 / 13540263240
Skill + Tool-Use0 / 1 / 15158403956
Fixed Failure RulesRule-Based3537253233
Unified ReflectorAdaptive P/S/U4449333546
Read-Only Structured MemoryRead-only P/S; U=04040293542
Shuffled Failure Context1 / 1 / 13744303436
Permuted Operator Weights1 / 1 / 15257394254
Exploration First Selection1 / 1 / 14551364647

Routing of failure-learning updates

RoutingGeneralFast ChargeLow TempHigh TempHigh Voltage
Random4345353647
Wrong3031192425
Skill Only4246273440
Prompt Only3232212828
Tool-Use Only3537202534

W1: Transferability across task regimes.

This is a good point.

  • We evaluate eight cross-task transfer directions by initializing each transfer task with frozen artifacts from a different control task. Transfer improves success rates by 5--14 percentage points over cold-start controls, showing that SAGE's distilled failure experience generalizes across task regimes.
  • We consider this a meaningful study and have included the results in the revised manuscript. The results are presented in Table 15 (pages 37--38).
ControlTransferIteration 2 ControlIteration 2 TransferGain (pp, ↑)
GeneralFast Charge5467+13
Fast ChargeGeneral4961+12
GeneralLow Temp.3044+14
Low Temp.General4958+9
Fast ChargeLow Temp.3037+7
Low Temp.Fast Charge5464+10
Low Temp.High Temp.3745+8
High VoltageHigh Temp.3742+5

W2 & Q3: Detailed protocol of physical evaluation.

We thank the reviewers for this critical point.

  • We have clarified the details of the physical evaluation protocol and the references for Task 5, and have included this discussion in the revised manuscript (Appendix A.10, pages 43--44).
  • We support our computational claims with effective and uniformly applied MD/DFT protocols, with explicit evaluation criteria. Our contribution is an agentic self-improvement framework that uses physical simulation feedback to narrow the search space and identify promising candidates, thereby reducing rather than replacing the high cost of wet-lab experimentation.

Q2: Study on guidance mechanism.

We agree with this point.

  • We conduct additional studies comparing different guidances under the same archive-based numerical optimizer, component pool, physical history, search operators, and evaluation budget. Fixed guidance uses the initial static design policy, rule-based guidance applies predefined failure-conditioned rules, and stateless reflection uses LLM-generated failure diagnoses and correct directions only within the current iteration, without persistent updates.
  • Because the archive, numerical optimizer, and evaluation budget are fixed, the consistent advantage of persistent, actionable policy updates shows that SAGE benefits not only from reacting to failures, but also from accumulating, routing, and operationalizing failure-specific experience to guide subsequent search under coupled constraints.
  • We have included these results in the revised manuscript, as shown in Table 12 (page 37).
GuidanceGeneralFast ChargeLow TempHigh TempHigh Voltage
Fixed Guidance2832212528
Rule-Based Guidance3536263134
Stateless Reflection3339263037
AuthorsOpenReview ↗

Kind Reminder: Review for Submission 11841

Dear Reviewer b1vL,

I hope this email finds you well. I am writing to gently follow up regarding the review for Submission 11841. We understand that you may have a busy schedule, and we simply wanted to check in, as we have not yet received your review.

If possible, we would be very grateful if you could kindly share an update on your progress or submit your review when you have a moment.

Thank you so much for your time, effort, and valuable support.

Best regards,

Authors of Submission 11841

AuthorsOpenReview ↗

Gentle Follow-Up on Review for Submission 11841

Dear Reviewer b1vL,

I hope you're well. A gentle follow-up on Submission 11841: the rebuttal period ends in about a day and we haven't yet received your feedback. We understand you're busy and just wanted to check in, so if possible, a brief update or your review would be greatly appreciated. Thank you for your time and support.

Best regards,

Authors of Submission 11841

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.