acceptodds
Withdrawn from ICLR 2027

RSI-Mol: Recursive Self-Improvement for Molecular Property Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Although molecular property prediction has advanced substantially through large-scale pretraining and sophisticated neural architectures, most existing approaches remain static, with limited ability to learn from prediction failures and refine the task-relevant knowledge and representations that guide subsequent predictions. We argue that molecular property prediction can instead be formulated as recursive self-improvement, where execution feedback is used not only prediction outcomes but also the molecular reasoning state used to generate subsequent predictions. To this end, we introduce RSI-Mol, a recursive self-improving framework for molecular property prediction that iteratively evolves its molecular reasoning state through interactions between large language models (LLMs) and downstream predictive models. RSI-Mol realizes recursive self-improvement through two coupled processes: in a self-improvement proposal process, LLMs reason over task-specific knowledge and the current molecular reasoning state to propose task-relevant physicochemical features and predictive substructures for downstream prediction; in a recursive feedback update process, prediction errors and SHAP-based attribution signals produced by the downstream predictor are fed back into the reasoning process to assess the current proposals and guide the refinement, elimination, and discovery of molecular features and substructures. The resulting updated molecular reasoning state, including task features, predictive substructures, feature pools, and failure-case context, is then reused as the starting point of the next reasoning round, where a new downstream predictor is constructed and evaluated. In this way, execution outcomes from one iteration directly update the reasoning state that governs the next, forming a recursive loop of proposal generation, prediction, diagnosis, and state refinement. Beyond improving predictive performance, RSI-Mol provides interpretable molecular evidence through attribution-guided feedback and chemically meaningful substructure refinement. Extensive experiments on eight MoleculeNet benchmarks demonstrate that RSI-Mol consistently outperforms strong unimodal and multimodal baselines, achieving an average AUC of 81.21% across single-task and multi-task settings. Code and implementation details are available at https://anonymous.4open.science/r/RSI-Mol-E767/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Recursive Self-Improvement, Molecular Property Prediction, LLMs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7AgaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes RSI-Mol, an LLM-assisted framework for molecular property prediction that iteratively updates an explicit molecular representation.

Critical strengths

The complete system shows promising performance across several benchmarks, supported by representation ablations.

Critical weaknesses
  1. The attached code is useless. No clear instructions to execute or understand the code.

  2. Writing. 2.1. What is "candidate" described in Sec. 3.1? Are they candidate features, candidate strategies, or candidate proposals?

  3. Methodology. 3.1. How do the authors ensure the effectiveness and robustness of your method? 3.2. I think this paper is not RSI work; it just uses LLMs to do automated feature selection or feature engineering (probably not even comprehensive enough to be called engineering). Novelty is very low and not interesting enough for ICLR.

  4. Experiments. 4.1. The experiments do not isolate the benefit of the recursive state updates. Section 3.4 identifies the inherited exclusion archive as the component distinguishing recursion from repeated execution, but there is no controlled ablation of this archive, the failure-case context, or attribution-guided pruning. 4.2. The evaluation does not establish the advantage over the closest alternatives. 4.3. These datasets are old datasets and highly likely to be involved in LLM training datasets; using them as benchmarks does not support the effectiveness.

Minor:

  1. Figure 2 needs a major revision. Many numbers are unclear.
Questions for authors

See Critical Weaknesses.

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

This paper proposes RSI-Mol, a recursive self-improvement framework for molecular property prediction. It combines LLM-generated molecular features and substructures with downstream prediction feedback to iteratively refine a persistent molecular reasoning state. Experiments on eight MoleculeNet datasets report competitive predictive performance, supported by feature attribution analyses, ablation studies, and an extension to drug–target interaction prediction.

Critical strengths
  1. RSI-Mol integrates LLM-based molecular reasoning with prediction errors and SHAP attribution to refine molecular representations. The persistent feature pool, exclusion archive, and failure-case context provide an explicit mechanism for accumulating task-specific knowledge.
  2. RSI-Mol achieves the highest reported average AUC across eight MoleculeNet benchmarks. The additional drug–target interaction experiment further demonstrates the potential applicability of the framework beyond standard molecular property prediction.
Critical weaknesses
  1. The existing ablations mainly examine feature components and model selection, without directly isolating the contribution of recursive state updates. The rollout analysis shows improved validation performance, but does not establish whether feedback-driven recursion provides benefits beyond iterative feature search. This limits the evidence supporting the paper's central contribution.
  2. The paper does not clearly establish whether all reported baselines use identical scaffold splits and evaluation settings. Moreover, LLM4SD, a closely related LLM-based molecular prediction approach discussed in the paper, is absent from the experimental comparison. These limitations weaken the evidence for the claimed advantages over existing methods.
Questions for authors
  1. Can the authors provide comparisons against a single-round variant and a budget-matched iterative variant without recursive feedback?
  2. Were all baseline results obtained using identical scaffold splits, random seeds, and task inclusion criteria?
  3. How were the SHAP-based pruning thresholds determined?
  4. How does the framework control potential overfitting from repeatedly using validation errors to guide feature refinement?
Reviewer_e13UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RSI-Mol that formulates molecular property prediction as a recursive self-improving system centered on a persistent molecular reasoning state. In each round, an LLM generates candidate task features and SMARTS substructures conditioned on the current state, RDKit validates them, a predictor is trained on the feature pool, and SHAP attributions on validation errors are used to prune weak features and provide context for the next round. This propose-execute-feedback loop achieves an average ROC-AUC of 81.21% on eight MoleculeNet datasets, outperforming UniMol and MV-Mol. The idea of using an LLM to generate features and iteratively refine them is interesting and well-motivated.

Critical strengths

The framework introduces a logical iterative loop where the LLM proposes hypotheses, RDKit verifies their chemical validity, and SHAP attributions along with the error archive serve as memory to guide subsequent rounds. This mechanism has practical value for molecular property prediction by combining LLM generation with rule-based validation. Empirically, the method is evaluated on a substantial number of datasets, and the reported average AUC outperforms most of the evaluated baselines.

Critical weaknesses

The mechanism for selecting failure cases remains unclear, and the robustness of anchoring on static top-k failure cases is not well justified. Furthermore, reusing the same validation set across rounds raises concerns about validation-set leakage and optimistic performance estimation. The paper also lacks sufficient reproducibility details regarding the feature pruning schedule and does not adequately justify why this iterative loop is preferable to simpler feature engineering pipelines.

Questions for authors
  1. How are ties handled when multiple failure cases are nearly indistinguishable, and why was a static top-k approach chosen over dynamic selection?
  2. Are invalid or weak features pruned dynamically in each round, and what is the exact pruning schedule?
  3. Could the authors report results under an independent validation protocol to rule out potential leakage?
  4. What are the specific reproducibility details (e.g., prompts, exact feature sets, hyperparameters)?

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

Related papers

Discussion (0)

Sign in to comment.