acceptodds
Under review as a conference paper at ICLR 2027

MARS: Reflective Tree Search for Molecular Structure Elucidation from Mass Spectra

Abstract

Mass spectrometry (MS) is a fundamental tool for molecular identification, yet automatically elucidating molecular structures from MS spectra remains a long-standing challenge. Existing approaches either rely on retrieval-based methods constrained by database coverage, or on one-shot *de novo* generation, which often yields chemically implausible structures due to the lack of verification mechanisms. Inspired by the iterative reasoning workflow of human experts, we propose **MARS**, a **M**olecular **A**gent framework for **R**eflective Tree **S**earch based on MS spectra. MARS reformulates one-shot molecular prediction as a structured, reflection-driven search process powered by LLMs, following a three-stage expert workflow of intuitive hypothesis, logical reflection, and local refinement. It first retrieves structurally similar references from a large-scale external knowledge base, and an Intuitive-Prior Chain-of-Thought (IP-CoT) synthesizes them with spectral evidence into high-quality initial candidates. A Spectral Diagnostic Chain-of-Thought (SD-CoT) then translates spectrum-level discrepancies into verbalized correction signals, which guide a Monte Carlo Tree Search (MCTS) to iteratively refine candidates within the chemical space. MARS achieves competitive performance on NPLIB1 and MolPuzzle datasets and generalizes well to real wastewater spectra under distribution shifts, highlighting the value of combining prior-guided hypotheses with spectrum-grounded structural refinement. Our code can be found at https://anonymous.4open.science/r/MARS-DC0B.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure elucidation, LLM agents, Tree search.

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_2PMkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MARS, an LLM-based framework for molecular structure elucidation from mass spectra. MARS combines retrieval-based initialization (IP-CoT), spectrum-grounded reflection (SD-CoT), and Monte Carlo Tree Search to iteratively refine candidate structures. Experiments on NPLIB1, MolPuzzle, and MassSpecGym show improvements over several existing baselines. The paper also includes component ablations, robustness analyses, and qualitative examples on real wastewater spectra.

Critical strengths
  1. The paper addresses a challenging and practically relevant problem. Reformulating molecular structure prediction as an iterative search process is well motivated. In particular, using spectral discrepancies to guide structural edits is a sensible approach to grounding LLM reasoning in external evidence.

  2. The experimental evaluation is reasonably comprehensive. The authors evaluate MARS on multiple benchmarks and provide component ablations, robustness analyses, and comparisons under shared candidate initialization. These experiments offer useful evidence for the benefits of iterative structural refinement.

Critical weaknesses
  • The comparison with existing methods is missing an important baseline. Table 1 compares MARS mainly against earlier methods such as DiffMS and MS-BART. However, FRIGID (Bohde et al., 2026) [1], which is already cited in the paper, reports substantially higher accuracy on NPLIB1 (25.03% Top-1 and 33.37% Top-10) than MARS (8.91% and 15.87%). Although differences in evaluation settings need to be considered, I would expect a direct comparison with the full FRIGID method, rather than only using it in the separate refinement experiment.

  • I am not fully convinced of the methodological novelty over existing work. K-MSE (Zhuang et al., 2025) [2] already combines external molecular knowledge, spectrum-based evaluation, and MCTS for structure elucidation. The main difference appears to be the specialized spectral-diagnostic prompting used to guide candidate edits. Moreover, Table 3 shows relatively modest improvements over FRIGID refinement (+1.52 Top-1 and +2.02 Top-10 points). I think the paper needs to establish more clearly what SD-CoT contributes beyond existing search and refinement mechanisms.

  • Some experimental comparisons are difficult to interpret. Several supervised baseline results in Table 1 are taken from previous work, and DiffMS is evaluated on a halogen-free subset of MolPuzzle. It is therefore unclear whether all reported comparisons use equivalent evaluation conditions. I also noticed that MARS achieves 60.00% Top-10 accuracy in Table 4 but 66.35% in Table 5. The difference in evaluation settings, if any, should be clarified.

  • The inference cost is quite substantial. MARS requires 372 seconds and approximately 20.6k tokens per molecule, considerably more than the supervised baselines. The reported 12.4 seconds with 30-way parallelization reflects improved throughput rather than lower single-sample latency. Given that MCTS accounts for much of the performance improvement, a compute-matched comparison would help assess whether the additional search cost is justified.

  • The evidence for real-world generalization remains limited. The wastewater experiment presents only three qualitative examples, without aggregate quantitative results. Meanwhile, MARS achieves just 2% Top-1 and 3% Top-10 accuracy on the full MassSpecGym evaluation. These results suggest that reliable identification of challenging unseen molecules remains difficult, and I would be cautious about drawing strong practical generalization conclusions from the current evidence.

[1] Bohde M, Liu H, Manjrekar M, et al. FRIGID: Scaling Diffusion-Based Molecular Generation from Mass Spectra at Training and Inference Time[J]. arXiv preprint arXiv:2604.16648, 2026.

[2] Zhuang X, Wu B, Cui J, et al. Boosting LLM’s molecular structure elucidation with knowledge enhanced tree search reasoning[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025: 22561-22576.

Questions for authors
  1. Could the authors compare the complete MARS framework against FRIGID under the same evaluation protocol on NPLIB1?

  2. How does MARS compare with K-MSE and FRIGID when controlling for the number of generated candidates, search iterations, and computational budget? Are the improvements in Table 3 statistically significant?

  3. Could the authors clarify the different results reported in Tables 4 and 5, including the evaluation subsets and search configurations?

  4. Can the authors provide quantitative evaluation on a larger set of real-world wastewater spectra to support the generalization claims?

  5. Citation formatting needs some attention. Several references appear in the textual citation format even when they are not grammatically integrated into the sentence (e.g., “Kind et al. (2018); Wen & Zhu (2015); Wang et al. (2016)” in the Introduction). These should instead use parenthetical citations, such as “(Kind et al., 2018; Wen & Zhu, 2015; Wang et al., 2016),” typically produced by \citep{} in LaTeX. Could the authors check and correct the citation formatting throughout the paper?

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

This paper presents MARS, an LLM-agent framework for molecular structure elucidation from tandem mass spectra given a target molecular formula. MARS retrieves fingerprint-similar molecular references, uses an intuitive-prior CoT for initialization, generates spectrum-discrepancy-based edits via a spectral diagnostic CoT, and explores candidate structures using a reward-guided MCTS procedure. Experiments evaluate the framework on NPLIB1, MolPuzzle, and selected MassSpecGym and wastewater cases.

Critical strengths
  1. The paper addresses a meaningful ML-for-science problem where exact structure identification is difficult and conventional retrieval is constrained by library coverage. The decomposition into retrieval-grounded initialization, spectral diagnosis, and search is intuitively well motivated.
  2. Separating the fingerprint-based retrieval pathway from the CFM-ID-based spectrum comparison makes the intended role of SD-CoT more concrete than a generic “LLM reflects on its answer” formulation. The explicit use of a forward simulator is a sensible attempt to anchor generation in external evidence.
  3. The empirical results are encouraging in several respects. DeepSeek-V3.2 MARS substantially exceeds its direct LLM-agent baselines on MolPuzzle, improving Top-10 accuracy from 60.49% for Thinking and 34.15% for Reflexion to 66.35%. The ablation in Table 5 also shows a substantial drop when MCTS is removed, from 66.35% to 36.59% Top-10 accuracy, which is consistent with search being an important component rather than an ornamental addition.
  4. The candidate-refinement experiment is a useful framing of MARS as a refinement method rather than only an end-to-end generator. Starting from FRIGID-base, MARS improves Top-1/Top-10 accuracy to 19.81%/24.02%, compared with 16.09%/18.19% without refinement. This is a practically relevant use case, since stronger pretrained molecular generators can provide better starting points than an LLM prompted from scratch.
  5. The authors make a serious effort to study robustness rather than reporting only one aggregate number. Figure 3(b-d) and Table 5 attempt to probe perception failures, reward weighting, retrieval quality, module removal, and search budget. This breadth is valuable, even though several analyses need stronger controls and clearer interpretation.
  6. Figure 4(a) is an effective qualitative illustration of the intended behavior under weak retrieval. It shows a candidate progressing from low similarity through several modified structures to an exact match, which is at least consistent with the claim that the system can occasionally recover from an imperfect initialization.
Critical weaknesses
  1. MCTS implementation. I found the search procedure a little hard to follow. For example, can an action involve reconstructing the whole molecule, or is it limited to local edits? How are new branches and duplicate candidates handled? Also, Equation (5) uses a fixed 0.5/0.5 backup rather than the usual visit-weighted update in MCTS. Some pseudocode and a comparison with standard backup strategies would help justify this choice.
  2. Reward and spectral diagnosis. Equation (4) does not clearly define the formula-consistency score. It is also unclear why this score is needed if the molecular formula is strictly enforced. The reward weights in Section 2.3 and Appendix Table 8 also appear inconsistent. For SD-CoT, I would like to know how peaks are matched and how the resulting differences are translated into structural edits. These details matter because spectral diagnosis is a central part of the method.
  3. Ablation experiments. Table 4 provides useful evidence for SD-CoT, but I could not find the exact evaluation subset or search budget. Why is MARS's Top-1 accuracy 27.00% here, compared with 25.86% in Tables 1 and 5? Clarifying the experimental setup would make this comparison easier to interpret.
  4. Baseline comparisons. The comparison with DiffMS in Table 1 may not be entirely fair, since DiffMS is evaluated on a halogen-free subset while the corresponding MARS result on that subset is not reported. I would also like to see a clearer comparison with K-MSE, which already combines reflection and tree search. Table 3 is useful, but it would help to explain more explicitly what MARS adds beyond K-MSE, ideally under matched experimental settings.
  5. Robustness and generalization. Some conclusions seem a little stronger than the evidence. For example, Top-1 accuracy drops from 25.86% to 20.90% on the perception-failure subset. The retrieval analysis also shows different trends for Top-1 and Top-10 accuracy. Reporting subset sizes and variability would help. Similarly, the three wastewater examples are interesting, but I would treat them as a qualitative case study rather than strong evidence of real-world generalization.
  6. Reproducibility and efficiency. Since the MolPuzzle spectra were manually transcribed, releasing the processed peak lists and explaining the transcription procedure would be helpful. MARS also has a fairly high inference cost. It would be useful to discuss the accuracy–cost tradeoff, since parallel processing improves throughput but does not reduce the computation required for each molecule.
Questions for authors
  1. Could the authors clarify how MCTS is implemented in practice, particularly how the LLM-generated edits form new branches and why Equation (5) uses a fixed 0.5/0.5 backup instead of the standard MCTS update?
  2. Since K-MSE also combines reflection and tree search for molecular structure elucidation, could the authors clarify what MARS adds beyond K-MSE? Would it be possible to compare the two methods under the same initialization and search budget?
Reviewer_mFXmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces an LLM agent framework for elucidating molecular structures from mass spectra. Using elaborate prompt design, a workflow, and a DB, it builds a systematic framework. The authors attempt to demonstrate its superiority through various experiments and comparisons.

Critical strengths
  • Using LLMs for molecular structure elucidation from MS seems to be a good research direction, and the agentic based approach is structured.
  • Extensive experiments were conducted.
  • A thorough ablation study was performed for each component of the proposed method.
Critical weaknesses
  • There is a lack of MS-specific technical contribution. It is a general molecular refinement using LLM prompts + MCTS. The actual processing of MS information is done by external tools (MIST, CFM-ID), and the LLM-side contribution is the prompt templates.
  • A deeper analysis is needed of what LLM diagnosis adds over forward-model-based refinement. Currently, Table 3 shows a performance gain of 1.52 pt, but such a performance difference without statistical significance is insufficient. In addition, the FRIGID row takes the performance reported in the original paper; was MARS evaluated on the same test set? Was it evaluated only on a subset?
  • Why is the FRIGID comparison missing from the main Table 1? It is essential.
  • An experiment comparing LLM-based scaffold extraction with deterministic tools is needed.
Questions for authors

None.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.