acceptodds
Under review as a conference paper at ICLR 2027

LLM-Guided Evolution under Search-Induced Distribution Shift for Organic Solar Cell Discovery

Abstract

Organic solar cell (OSC) discovery requires searching an enormous chemical space using only sparse experimental supervision. Efficient discovery requires exploring beyond the limited set of experimentally observed molecules, while optimization progressively concentrates candidates in the sparsely sampled high-performance tail, where conventional predictors become less reliable. To address these challenges, we introduce a closed-loop molecular discovery framework that couples feedback-conditioned LLM evolutionary search with a discovery-oriented high-tail evaluator. The LLM guides semantic mutations over experimentally grounded molecules, enabling adaptive structural exploration beyond fixed recombination heuristics while preserving chemically meaningful scaffolds. For candidate evaluation, we propose Aggressive Tail Fusion, which combines a data-driven model with an experimentally calibrated physics model to improve reliability in the high-PCE regime. It reduces MAE on the experimentally observed top-5% subset from to .

Primary area:
generative models
Keywords:
Organic Solar Cells, Molecular Discovery, LLM-Guided Evolution, Distribution Shift, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_bfhWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on organic solar cell discovery using LLM guided evolution. The two main claimed contributions of the paper are discovery-oriented evaluator for the high-PCE regime and Feedback-conditioned LLM evolutionary search. The paper also studies distribution shift during the search.

Critical strengths

The direction of the work presented by the paper is good and beneficial to domain specific researchers.

Critical weaknesses
  1. I do not believe the work is fit for an ML conference. The 3 claimed contributions are an open research direction but i am not convinced that the work address any of those.
  2. The work is highly domain specific and narrow. This does not take away from the soundness of the work but that it is more relevant for domain specific journal than ML conference.
  3. Nothing in the paper makes an attempt to understand distribution shift during the search. It is an existing phenonmenon that is studied across multiple papers. This work simply claims it and moves on. No information is presented on how actual prediction error changes under that shift or demonstrate that their method mitigates it.
  4. The contribution of better evaluation is not backed as comparison is with OOD samples but we have no way to verify if these are actually valid. Unfortunately this can only be claimed by real experiments or high fidelity simulations. I acknowledge most people dont have access to these resources however without real experimental evidence the conclusions are misleading. Their experiments supports better high tail accuracy, but is not equivalent to evaluation of OOD molecules. Without independent ground truth, i am not convinced of the reliability.
  5. Nothing in the LLM guided evolutionary search is even incremental to previous approaches. Their implementation is constrained selection with standard graph-level operators.
Questions for authors

See weaknesses.

Reviewer_46eVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an LLM-guided evolutionary framework for discovering Y6-type non-fullerene acceptors for organic solar cells. The method uses an LLM to guide fragment-based molecular edits and incorporates historical evaluation feedback into subsequent search decisions. It also introduces an Aggressive Tail Fusion evaluator that combines a data-driven predictor with a calibrated Scharber-style model, aiming to improve evaluation reliability under search-induced distribution shift, particularly in sparsely sampled high-performance regions. The authors evaluate the framework through property prediction on experimental OSC data, high-PCE subset analyses, and surrogate-guided molecular search.

Critical strengths
  1. The paper addresses an important practical challenge: maintaining evaluator reliability when molecular optimization moves toward sparsely sampled, high-performance regions.

  2. The curation of 746 experimentally characterized Y6-type acceptors provides useful data for OSC property prediction and molecular design.

Critical weaknesses
  1. The LLM guided evolutionary search lacks novelty.

The method uses an LLM to select predefined core, end-group, and side-chain fragments, constructs molecules through fixed genetic operators, and incorporates historical evaluation feedback to guide subsequent edits. These core design principles have clear precedents. MoLLEO [1] introduced LLMs into molecular evolutionary algorithms as mutation and crossover operators. AutoLeadDesign [2] combined fragment libraries, LLM-based molecular generation, and evaluation-driven iterative updates. ToolMol [5] integrated evolutionary search with LLM-guided structural modifications executed through RDKit-based editing tools, while ReDF [6] used retrieved examples and domain feedback to guide iterative molecular refinement. PROBE [3] integrated a prebuilt fragment library, LLM-guided local editing, predefined graph operations, and historical feedback. OSCAgent [4] has also explored LLM-based OSC molecular discovery.

The paper does not identify a substantive algorithmic innovation beyond these established mechanisms. The LLM primarily selects or ranks candidates within a predefined fragment space, molecular construction relies on prescribed structural operations, and historical feedback is supplied through the context of subsequent calls. The additions mainly consist of defining a fragment space for Y6-type acceptors and integrating existing editing, feedback, and evolutionary selection mechanisms. As presented, this constitutes a domain-specific adaptation of established approaches, without a distinct new search principle or adaptation mechanism. The claimed LLM search contribution therefore lacks methodological novelty.

[1] Efficient Evolutionary Search Over Chemical Space with Large Language Models. ICLR, 2025. arXiv:2406.16976.

[2] A Collaborative Framework Integrating Large Language Model and Chemical Fragment Space: Mutual Inspiration for Lead Design. arXiv preprint, 2025. arXiv:2507.13580.

[3] Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design. arXiv preprint, 2026. arXiv:2606.00555.

[4] OSCAgent: Accelerating the Discovery of Organic Solar Cells with LLM Agents. arXiv preprint, 2026. arXiv:2602.04510.

[5] ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

[6] Utilizing Large Language Models in An Iterative Paradigm with Domain Feedback for Molecule Optimization

  1. Gains over the matched baseline are limited, and evaluator accuracy and independent validation are insufficient to establish improvements in actual materials discovery.

Compared with Rule-based EA, which uses the same genetic operators, screening criteria, and evaluator, the highest Best PCE increases only from 19.4425% to 19.5813%, an absolute gain of approximately 0.139 percentage points. The largest improvement in AUC Top-10 is 0.0913. Meanwhile, the fused evaluator used during search achieves an overall R² of 0.683 and an MAE of 2.025 percentage points on the experimental dataset, indicating appreciable residual prediction error. The reported results do not establish that these small differences in surrogate scores translate into reliable candidate rankings or improved device performance.

Both search guidance and final assessment rely on the same fused surrogate evaluator, and the generated candidates lack independent performance validation. In particular, the paper provides neither first-principles checks of key electronic or optical properties nor experimental validation through synthesis and device testing. The experiments therefore do not rule out the possibility that the search exploits evaluator biases to obtain higher scores. Given the limited score improvements and the evaluator’s prediction errors, the evidence supports optimization of a particular surrogate objective but does not establish superiority in actual OSC materials discovery or more reliable identification of high-performance candidates.

  1. The independent contribution of historical feedback is not established because the required ablation is missing.

The paper summarizes performance changes associated with previous edits and whether the resulting candidates survive selection, then provides this information to the LLM as an important mechanism for improving subsequent decisions. However, Rule-based EA removes both LLM guidance and the LLM’s use of historical information, so the resulting performance differences cannot be attributed specifically to historical feedback. The paper lacks an ablation that removes only historical feedback while retaining the same LLM, current candidate information, and evolutionary settings. Consequently, it cannot distinguish improvements arising from the LLM’s prior chemical knowledge, current candidate information, or adaptation based on historical feedback.

An upward evolutionary trajectory alone does not demonstrate that the feedback mechanism is effective: repeated selection of high-scoring candidates and elite retention can themselves produce similar trajectories. The existing experiments do not isolate the effect of historical feedback and therefore do not substantiate its claimed independent contribution to search performance.

Questions for authors

Are the MAE improvements achieved by the physics-based branch and fused evaluator on the Top-5% subset statistically robust?** This subset contains only 76 device samples, but the number of distinct acceptor molecules represented and the sample counts in each test fold are unclear. Are the reported advantages sensitive to a few influential samples or to the particular data split?

Reviewer_36FxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies LLM-guided evolutionary search: an LLM uses previous search outcomes to guide fragment selection, while genetic operators construct new candidates. It also introduces a fused predictor to score candidates. My main concern is whether LLM guidance provides a reliable and meaningful improvement over the matched rule-based evolutionary baseline.

Critical strengths
  1. Rule-based EA uses the same evolutionary operators and evaluator, providing a relevant comparison for assessing LLM guidance.
  2. The paper evaluates the predictor against experimental data and examines search trajectories across several LLM backbones.
Critical weaknesses
  1. The reliability of the improvements is unclear. In Table 2, Best PCE increases from 19.4425 for Rule-based EA to 19.5813 for the strongest LLM result, approximately 0.14 percentage points. Other search metrics also improve, but uncertainty estimates are missing. Appendix C.2 repeats only DeepSeek, without corresponding repeated runs for Rule-based EA or the stronger LLMs. The statistical significance of the comparative gains is therefore not established.

  2. Higher predicted scores do not establish better molecular performance. All methods optimize and are evaluated by the same predictor. Its validation on existing experimental data is useful, but does not establish reliability on newly generated molecules.

  3. The contribution is difficult to assess from the presentation. The paper combines a new evaluator with LLM-guided search, but their individual contributions are not clearly separated. Sections 4.3–4.4 describe score improvements without sufficiently explaining what decisions LLM guidance improves over Rule-based EA or whether the gains justify its additional cost.

Questions for authors
  1. Can you report repeated-run means and uncertainty for Rule-based EA and the LLM variants under matched search budgets, and assess whether their differences are statistically reliable?

  2. What evidence shows that higher scores correspond to better molecular candidates rather than evaluator error under distribution shift?

  3. What specific search decisions does LLM guidance improve, and what additional API/token cost and runtime does it require?

Reviewer_MADNOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an LLM-guided evolutionary framework for organic solar cell discovery. It combines feedback-conditioned molecular fragment selection with Aggressive Tail Fusion, which integrates data-driven and physics-based PCE predictors. The framework aims to improve molecular optimization under search-induced distribution shift.

Critical strengths

The paper addresses an important problem concerning the reliability of surrogate evaluators during molecular optimization, particularly when optimization moves beyond the training distribution.

The proposed fusion improves prediction accuracy on experimentally observed high-PCE samples.

The experiments include multiple evolutionary baselines and LLM backbones, providing useful comparisons of different search strategies.

Critical weaknesses
  1. Limited methodological novelty. Aggressive Tail Fusion is essentially an asymmetric heuristic combination of two predictors, while the LLM mainly selects fragments for predefined genetic operators. The algorithmic contributions are limited.

  2. Search-induced distribution shift remains unresolved. The evaluator is validated on high-PCE subsets of existing experimental data, not on genuinely out-of-distribution molecules generated through optimization. The central scientific claim therefore lacks direct evidence.

  3. Prediction accuracy does not necessarily imply optimization effectiveness. Evolutionary optimization primarily depends on candidate rankings rather than absolute prediction accuracy. The standalone physics model even achieves a higher Top-5% Spearman correlation than the fusion model (0.378 vs. 0.329).

  4. Insufficient ablation studies. The paper lacks optimization-level ablations comparing data-driven-only, physics-only, and fused evaluators. It also fails to isolate the contribution of LLM-guided fragment selection against simpler adaptive selection strategies.

  5. Limited generalizability. The asymmetric fusion is specifically designed around the underestimation bias of the current data-driven predictor. Would it remain effective with different prediction models or reversed error patterns? The paper provides no evidence of such generalizability.

  6. Optimization results lack independent validation. All generated molecules are evaluated using the authors' own surrogate. The best predicted PCE improves only from 19.4425% to 19.5813% over Rule-based EA, without independent experimental or high-fidelity computational validation.

  7. Questionable synthesizability. The generated molecules have average SA scores of approximately 7.1–7.6, raising concerns about their practical synthesizability.

Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.