acceptodds
Withdrawn from ICLR 2027

EvoMol: Semantic Gradient Evolution for Rule-Guided Molecular Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) have shown strong potential for molecular reasoning, but their performance depends heavily on the reasoning context, and recurring errors are rarely retained as reusable strategies. We present EvoMol, a semantic gradient evolution framework that adapts the reasoning logic of frozen LLMs without updating model parameters. EvoMol represents the task context as a structured logic kernel containing task decomposition strategies, molecular constraints, and mechanistic heuristics. At each generation, EvoMol diagnoses model failures and organizes them as Actionable Side Information, including the observed error, underlying chemical principle, corrective action, and diagnostic confidence. This structured feedback guides rule-level mutations of the logic kernel. A multi-frontier Pareto consensus mechanism then selects candidates that perform consistently across instances and objectives. Across three molecular benchmarks and Qwen2.5 models from 3B to 14B parameters, EvoMol achieves the best performance in 49 of 60 settings under limited supervision and outperforms direct prompting and GRPO in most cases. Controlled analyses show that the evolved kernels capture task-relevant strategies beyond generic or irrelevant prompt content and transfer across mechanistically related tasks. EvoMol also improves multi-objective molecular optimization under competing property constraints. These results demonstrate that EvoMol provides a data-efficient and inspectable approach to adapting molecular reasoning with frozen LLMs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular optimization, drug 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_zu6oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EvoMol introduces a framework that continuously refines the reasoning strategies of LLMs by transforming failures in molecular reasoning into explicit, reusable rules, without updating model parameters. The authors represent the model’s reasoning context as a structured “logic kernel” containing task decomposition strategies, chemical constraints, and output verification rules, which is progressively refined through iterative evolution. EvoMol follows a reasoning–diagnosis–evolution–consensus selection pipeline. The paper refers to information gathered during task execution, such as observed errors and the associated chemical or reasoning principles, as “semantic gradients,” which provide directional feedback for prompt refinement. The framework also introduces a Reflective Proposer and uses a Multi-Frontier Pareto Consensus mechanism to iteratively generate and select the next generation of reasoning contexts. Experiments mainly demonstrate performance improvements on the Qwen2.5-3B, Qwen2.5-7B, and Qwen2.5-14B models.

Critical strengths

EvoMol transforms lessons learned from failures into inspectable, reusable reasoning rules and provides experimental evidence supporting their effectiveness. Its main contribution lies in consolidating failure feedback into explicit reasoning strategies.

Critical weaknesses
  1. The lack of comparisons with directly relevant prompt optimization methods leaves the method’s incremental value insufficiently established. Although the paper discusses feedback-driven methods such as GEPA[1] in the related work section, its main experiments compare EvoMol only with zero-shot direct prompting and GRPO, without direct comparisons against other prompt optimization methods. The reported results only show that EvoMol outperforms the static prompts used in the experiments and the particular GRPO configuration under a specific low-resource setting. They do not sufficiently establish the method’s superiority more broadly. In addition, the paper does not provide separate ablation studies to isolate the contributions of the structured semantic gradients and the Multi-Frontier Pareto Consensus mechanism.

  2. Lack of novelty. The core framework overlaps substantially with GEPA, and the idea of accumulating chemical reasoning experience as reusable knowledge is not new, as it has already been explored in ChemAgent[2]. Again, the paper should at least include a direct comparison with GEPA.

[1] Agrawal, L. A., Tan, S., Soylu, D., et al. (2026). GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. In Proceedings of the 14th International Conference on Learning Representations (ICLR 2026). [2] Tang, X., Hu, T., Ye, M., et al. (2025). ChemAgent: Self-updating Memories in Large Language Models Improves Chemical Reasoning. In Proceedings of the 13th International Conference on Learning Representations (ICLR 2025).

Questions for authors

Suggestions for future work:

Need additional experiments: under the same chemical tools, knowledge base, reflection model, and evaluation budget, does the proposed structured feedback significantly outperform simply providing raw error information to GEPA?

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

EvoMol evolves textual rules for molecular tasks while keeping model weights frozen. The system structures failure feedback into ASI, comprising errors, chemical principles, corrective actions, and confidence information. A knowledge-assisted reflection module proposes rule revisions, after which mini-batch screening and multi-front Pareto consensus selection determine which candidates to retain. The paper compares Direct, GRPO, and EvoMol using Qwen2.5 models with 3B, 7B, and 14B parameters across multiple types of molecular tasks. It also reports results on rule transfer, controlled experiments on rule content, and multi-objective optimization. Across the 20 tasks and three model sizes shown in Table 1, EvoMol achieves the best mean performance in 49 of the 60 cells. This does not imply 49 statistically significant wins, nor can it automatically be taken as representative of all 51 tasks described in the paper.

Critical strengths
  1. The problem is practically relevant, and the outputs are inspectable. Learning reusable chemical constraints with a frozen model is more informative for diagnosing recurring errors than presenting only a final score. The explicit ASI structure and kernel-evolution process are reasonably clear. Figure 11 shows a progression from representation validity and structural features to chemical class/scaffold strategies, providing observable intermediate artifacts (page 4, Equations (4)–(8); page 19, Figure 11). This supports inspectability, but does not establish the chemical correctness of every induced rule.

  2. Some improvements are large and span more than one task type. In Table 1, 3B ring-count MAE decreases from 1.450 for Direct and 1.050 for GRPO to 0.467. MolInstruction 3B retrosynthesis improves from 0.183 and 0.146 to 0.790; 7B molecular design improves from 0.071 and 0.024 to 0.438 (page 6, Table 1). These are author-reported means, not independently reproduced results, but they justify further study. Table 1 also retains settings where a baseline wins rather than implying universal superiority.

  3. Meaningful content controls and transfer analyses are already present. Figure 2 compares not only the initial prompt but also Generic, Nonsense, and Cross-task conditions. For 3B forward reaction, Full scores 0.6125, Generic 0.1787, and Nonsense 0.1128; for 14B retrosynthesis, Cross scores 0.7614, close to Full's 0.7955 (pages 7–8, §4.3 and Figure 2; pages 16–19, D.4–D.5). This supports the usefulness of task-relevant prompt content and shows that transfer varies across tasks. The causal interpretation nevertheless needs to remain narrower, as discussed below.

  4. Multi-objective results broaden the empirical scope. For 7B, GSK3B + QED hypervolume (HV) is 0.5236 versus 0.2151 for Direct; for 3B, DRD2 + QED + LogP improves from 0.0209 to 0.5037 (pages 8–9, Figure 3 and §4.4). Appendix D.7 provides information about 50 source molecules per task, eight generations of search, normalization, and the HV reference point (page 20). These results concern computational proxy objectives, not demonstrated drug efficacy or laboratory validation. The ethics statement also acknowledges this boundary (page 10).

Critical weaknesses

1. The closest text optimization baselines are missing, making it difficult to assess the main novelty and the sources of the gains. — The primary issue driving the recommendation. Pages 2–3 do not adequately explain how the method differs from ProTeGi and GEPA. ProTeGi already uses textual gradients derived from errors, together with search; GEPA already incorporates reflection based on execution traces, a pool of historical candidates, mini-batch screening, and parent selection based on instance-level Pareto support. These are not simple optimizers that rely solely on scalar feedback.

2. The main text and the algorithm describe the candidate pool and generational updates inconsistently, making it difficult to reconstruct the actual method. — Confirmed textual inconsistency; the actual implementation remains to be clarified. Equation (9) (p. 4) describes the candidate set as mutant candidates generated from the current population, and Equation (14) (p. 5) selects the next generation from that set. Algorithm 1 (p. 13), however, initializes a cumulative candidate pool, selects parents from the historical pool, adds newly generated candidates to it, and makes the final selection based on support. The two descriptions differ in whether historical candidates are continuously retained and in the pool from which selections are made.

3. “Improvements in chemical reasoning” have not been adequately separated from formatting corrections and learning task conventions. — An alternative explanation supported by evidence, not a confirmed cause. The Generic and Nonsense controls on page 17 each consist of four sentences, but do not adequately match the final rules in length, output contract, and formatting information. The early changes shown in Figure 11 (p. 19) include SELFIES syntax repairs, and the examples in Appendix D.8 (p. 20) also include output format corrections.

Questions for authors

With the reflection model, knowledge access, and feedback budget held constant, how much of an advantage does EvoMol retain over GEPA-style search? The answer would directly affect the assessment of its novelty and empirical contribution.

Reviewer_5atPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EvoMol, a framework that iteratively optimizes task prompts for frozen language models through structured error diagnosis, chemistry-informed rule revision, and multi-front Pareto selection. Across multiple molecular tasks and model scales, EvoMol outperforms direct prompting and the GRPO baseline evaluated in the paper under low-resource settings, while also demonstrating potential for multi-objective molecular optimization.

Critical strengths

The paper addresses a practically relevant problem. It focuses on adapting language models to molecular tasks with very limited labeled data, improving performance by updating prompts rather than model parameters. This is relevant to applications where labeled data are scarce or model fine-tuning is impractical.

The method delivers broad performance improvements under the low-resource settings evaluated in the paper. EvoMol achieves the highest mean performance in 49 of the 60 model–task configurations reported in the main results table. Substantial gains on certain tasks, such as ring counting and molecular generation on MolInstruction, support the effectiveness of the overall framework under the tested conditions.

The paper considers trade-offs among multiple molecular properties. It applies the method to the joint optimization of multiple property objectives and evaluates the resulting solution sets using hypervolume (HV), extending the experimental scope beyond single-property optimization. The results demonstrate the method’s potential for multi-objective molecular optimization.

Critical weaknesses

The distinction from GEPA is unclear. EvoMol shares several key components with GEPA, including feedback-driven reflection, prompt revision, and Pareto-based selection. The paper should better acknowledge GEPA’s ability to accumulate rules across examples and clarify what its chemistry-specific feedback and multi-front selection add.

Direct comparisons with related prompt optimization methods are missing. The experiments compare only against Direct Prompting and GRPO. A comparison with GEPA under matched tasks, reflection models, domain knowledge, and budgets is needed to establish EvoMol’s relative advantages.

The dependence on external knowledge remains unclear. The experiments do not show whether EvoMol mainly draws on existing chemical knowledge or can develop effective strategies when that knowledge is incomplete or incorrect. Knowledge removal, incorrect-rule injection, and tests on unseen mechanisms would help clarify this limitation.

Questions for authors

see the weaknesses

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

Related papers

Discussion (0)

Sign in to comment.