acceptodds
Under review as a conference paper at ICLR 2027

Think in Fragments: Fragment-Centric Reasoning and Credit Assignment for LLM Molecular Optimization

11 views

Abstract

Large language models (LLMs) provide a flexible interface for molecular optimization, yet most LLM-based methods describe intended changes using chemical concepts while generating molecules token by token and learning from a single molecule-level reward. This granularity mismatch makes it difficult to faithfully execute chemically meaningful edits and assign credit to the structural changes responsible for property improvement. We propose , a fragment-centric framework that uses functional groups and molecular scaffolds as a common interface for reasoning, editing, experience reuse, and policy optimization. combines hierarchical retrieval of task-relevant fragments, an evolvable dual-trace memory of optimization experience, and fragment-aware group relative policy optimization (FA-GRPO), which estimates the contributions of edited fragments and propagates fragment-level advantages to their corresponding output tokens. Across the 23 tasks of the Practical Molecular Optimization benchmark under a 1,000-query budget, achieves higher aggregate Top-10 AUC than the strongest evaluated LLM baseline across the tested backbone models. These results demonstrate the effectiveness of aligning molecular reasoning and learning with chemically meaningful fragments for improving the sample efficiency of LLM-based molecular optimization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM, RL, Frag,Molecular Optimization,Agent

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_sky8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on using a fragment-based method for LLM-based molecular optimization. The LLM reasons about edits at the level of functional groups, but generates molecules token by token, and the generated molecule is evaluated by an oracle as a whole for a single reward. The paper uses MacFrag to build a fragment library and stores the editing history in a structured memory. Rather than using only the final reward after editing, the paper uses Monte Carlo Shapley values to estimate the contribution of each fragment edit and assigns token-level credit accordingly. Experiments are conducted on the PMO benchmark.

Critical strengths
  1. The paper focuses on the credit assignment problem in molecular editing. Since functional groups are closely related to specific chemical properties, assigning credit to fragment edits rather than to the whole molecule may improve GRPO training and the resulting policy.
Critical weaknesses
  1. The motivation is not well verified. The paper claims a mismatch among reasoning at the functional-group level, generation at the token level, and rewards at the whole-molecule level. However, Figure 1 only measures whether generated molecules follow the intended edit strategy, which is at most the gap between reasoning and generation. This is a different problem and cannot support the paper's claim.

  2. The proposed credit assignment is not clearly connected to the claimed mismatch. Assigning the same advantage to all tokens is a general credit-assignment issue in GRPO, which would exist even without the mismatch, and the paper does not show that standard GRPO fails because of it. Moreover, the benefit of FA-GRPO is only weakly supported by Table 2, which uses a single backbone and reports no variance.

Questions for authors
  1. Compared with common fragment decomposition methods such as BRICS, what is the advantage of MacFrag, and why is it chosen?

  2. Why does FragThinker (GPT-4) perform better than MOLLEO? Do both use the same LLM backbone and initial molecule pool?

  3. Does the model generate an explicit reasoning process before the edit? If so, Eq. 7 appears to assign the molecule-level advantage to all reasoning tokens. Does this mean the reasoning that motivates an edit receives no fragment-level credit?

Reviewer_9urdOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FragThinker, a framework for LLM-based molecular optimization that combines fragment retrieval, an experience memory, and fragment-aware GRPO (FA-GRPO). Its central idea is to estimate the contributions of molecular edits and assign corresponding learning signals to the output tokens. The approach is plausible, but the current evidence does not clearly establish its novelty or demonstrate improved sample efficiency under comparable total oracle costs.

Critical strengths
  1. Using fragments as a shared interface for retrieval, editing, memory, and learning is a coherent design motivated by chemically meaningful interventions.

  2. Evaluation across 23 PMO tasks, two LLM backbones, and component ablations provides a useful starting point for assessing the framework.

Critical weaknesses
  1. The contribution is insufficiently distinguished from existing work on fragment-based optimization, retrieval-guided editing, substructure attribution, and fine-grained policy optimization.
  2. The central claim—that fragment-level credit improves learning and sample efficiency—remains inadequately supported because attribution requires additional oracle evaluations and the controls do not isolate its specific benefit.
Questions for authors
  1. Does the advantage persist at equal total oracle cost? Target-task evaluations used for training rollouts, counterfactuals, and Shapley estimation are excluded from the reported search budget. These evaluations still provide information that can improve subsequent proposals. Please compare FA-GRPO with standard GRPO under the same total target-task oracle budget and clarify when policy updates occur relative to search.

  2. What is new relative to closely related molecular methods? Please discuss and distinguish FaST (2111.01009) and FREED(2110.01219) for fragment-based RL, ChatDrug (2305.18090) for retrieval and feedback, SME(s41467-023-38192-3) for substructure attribution, and FORGE(2605.10230) for context-aware fragment ranking, replacement, and experience reuse. Which capabilities are technically new, and which are integrations of existing ideas? Direct comparisons with the closest methods would strengthen the novelty claim.

  3. When does fragment attribution provide additional information? With a shared parent, one edit per candidate, and an oracle score as reward, normalized Shapley contributions largely reduce to the standard molecular advantage, apart from clipping. Please report the distribution of edits per candidate and compare against simple edit reversion, shuffled credit, and edit-token weighting without attribution. Uncertainty estimates for these comparisons are also needed.

  4. Is the proposed reasoning–execution gap directly demonstrated? Figure 1 does not consistently favor the proposed method, and the consistency metric lacks sufficient evaluation detail. Please specify its definition, sample size, and assessment procedure, and show that improving edit fidelity explains optimization gains. Reproducibility also requires the retrieval-ranking procedure, SFT data sources and splits, and the online update schedule.

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

This paper argues that existing methods of using LLM for molecule optimization face the issue of fragment-level reasoning-execution-credit gap. While the LLMs reason the molecular edits on fragment level, the execution is on token level (e.g. SMILES), and the final evaluation is on the entire molecule level. To address this limitation, the authors propose FragThinker, which implements fragment-level edits and can attribute reward directly to the fragments as well. Benchmark on PMO shows that FragThinker presents superior results, where the FragThinker-Qwen3 scores first place in 20/23 tasks. Ablation study further demonstrates the three key components of FragThinker (HFKR, EDTM, FA-GRPO) all contribute to the improved performance.

Critical strengths
  • High novelty: this paper proposed a quite novel method to attribute fragment-level reward to smiles tokens, solving the problem that many chemical fragments are discontinuous in the smiles strings due to the existence of branches and rings. To me, this is the main roadblock before why LLM-based molecular editing cannot easily do the fragment-based effect evaluation, which is quite mature in graph-based methods.
  • Comprehensive benchmark and ablation design: the authors have performed a very solid benchmark versus existing methods to demonstrate the advantage of FragThinker. Ablation study also clearly demonstrates the contribution of different components in FragThinker.
Critical weaknesses
  • The paper is not well-written: I found the current manuscript is relatively difficult to follow. It proposed several new concepts but did not provide enough details. Examples include using HFKR and EDTM without defining it (line 83-84), presenting results without experiments setting up and interpretation (Figure 1), large but sparse figures with limited information (Figure 2 and 4). I think it can benefit a lot from polishing the writing, removing unnecessary graphics, and using the saved space to add more discussions.
Questions for authors
  1. In Table 1, the authors show that the full model outperforms the sequence-level GRPO in top-10 AUC. However, Table 4 suggests that the full system has 2452 oracle calls, which is ~25% higher than the sequence-level GRPO (1840). How could we know that the better performance of the full model does not originate from more access to Oracle?
  2. The author said that RL Oracle calls are excluded from the search budget in benchmark tests. Would it be more reasonable to use the total Oracle calls as the budget for a fair comparison with other methods? I understand that B=1000 is a typical budget for PMO which might be too small for tuning via the FragThinker framework, but I think the authors could increase it for other methods to match the needs of FragThinker. This can also help evaluate the effect of oracle calls in model performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.