acceptodds
Under review as a conference paper at ICLR 2027

MolGoT: Organizing LLM Molecular Reasoning with Structure-Guided Graphs of Thought

Abstract

Large language models (LLMs) are increasingly used to understand, modify, and generate molecules through natural-language instructions. Recent approaches expose atom identities and bond relations in explicit molecular representations, reducing the need to reconstruct connectivity from compact string syntax before performing chemical operations. However, making structure accessible does not determine how intermediate decisions should be coordinated: observations from connected regions may depend on one another, while separately proposed edits can interact when combined. Graph of Thoughts offers a framework for this coordination by representing intermediate results as nodes and their dependencies as edges, enabling findings from different reasoning paths to be combined and revised. Molecular structure can ground these operations by identifying which atoms and regions each state concerns and where their consequences may interact. Building on this connection, we introduce MolGoT, a training-free method that retains SMILES and organizes molecular reasoning through structure-guided graphs of thought. MolGoT constructs a shared structural context and maintains local evidence and executable molecular candidates, with task-specific evaluation, integration, and revision for molecular understanding, analysis, editing, and generation. Across eleven molecular tasks, MolGoT achieves the highest average accuracy among the four evaluated pipelines. These results suggest that structure-guided thought graphs provide a controllable framework for improving both the accuracy and token efficiency of molecular reasoning in LLMs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular reasoning, graph of thoughts, molecular subgraphs, large language models, test-time computation

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_Cy6oOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work introduces a novel scoring scheme that integrates the Graph of Thoughts (GoT) framework into molecular editing and understanding tasks, enabling more robust agentic workflows with fewer errors and improved token efficiency. The study is thorough and includes extensive ablation experiments, demonstrating that, for tasks such as organic molecule editing, the performance improvements primarily arise from the combined effect of task orchestration, scoring, and merging strategies. Each individual component, when applied in isolation, does not necessarily lead to substantial improvements.

Critical strengths
  1. The conventional GoT framework is relatively abstract and does not use specific physical or chemical meanings. The operations are branching/merging between different reasoning states. This work merged molecular editing information and domain-specific knowledge into the framework, making the reasoning process on the graph more controllable and better guided for chemical tasks. This integration also enhances interpretability and reduces the generation of invalid molecules.

  2. The experimental evaluation is comprehensive, with extensive ablation studies and detailed analysis. The authors objectively describe the contributions and even the limitations of individual components, providing a convincing evaluation of the proposed framework.

Critical weaknesses
  1. Lack of tests in larger molecules (molecule weight's influnence on the results). In this work, the authors use relatively small molecules which can be separated into three regions for testing. I was wondering about the robustness of the current scoring functions towards editing larger systems. Since there are some similarity checks in the scoring, and a larger system could reduce the distinguishability of the scores.

  2. The ablation results presented in Fig. 5a are particularly interesting. Among the evaluated components, count route appears to have the largest impact on performance. Interestingly, removing certain components, such as repair, seems to improve performance for some models, while other intuitively important components, such as merge selection, show relatively limited effects. These observations deserve a more detailed analysis to understand which aspects of the reasoning process are affected by each component and why such effects arise. Furthermore, these results may suggest that the current ablation study does not fully disentangle the contributions and interactions of different components. If feasible, the authors are encouraged to extend the ablation study to include additional components or combinations of components, which could provide a clearer understanding of their individual and synergistic effects.

Questions for authors
  1. How does the proposed scoring function accommodate molecular editing tasks that require substantial modifications to the original molecule? In Eq. (1), the similarity term SS is defined based on the similarity to the original molecule. While this term may help prevent unreasonable deviations from the starting structure, it could potentially become counterproductive when substantial structural changes are necessary to achieve the desired outcome. In such cases, could a more accurate and desirable edited molecule actually receive a lower score due to its reduced similarity to the original molecule? It would be helpful if the authors could clarify how the scoring function handles such scenarios and whether this introduces a bias toward conservative molecular edits.

  2. It would be better to add some more analysis and tests in ablation studies if possible.

6 comments

W1. Tests on larger molecules and the robustness of the scoring function.

We thank the reviewer for this suggestion. We ran the editing tasks on a new set of molecules of up to 3819 g/mol. MolGoT is more accurate than CoT at every molecular size, the region decomposition still helps on the largest molecules, and the similarity scores of candidates do not become harder to distinguish as molecules grow.

Setup. 120 distinct molecules with 9–169 carbon atoms (201–3819 g/mol), split by carbon count into four groups of 30. Each group has 10 addition, 10 deletion, and 10 substitution tasks. The model is Qwen3-8B. All numbers are accuracy in %.

(1) MolGoT improves over CoT in every size group. The table compares SMILES-CoT and MolGoT in each group. The last column is MolGoT with the whole molecule treated as one region, which we discuss in (2).

Carbon atomsMolecular weight (g/mol)SMILES-CoTMolGoTMolGoT, single region
9–27201–57130.058.340.0
28–52497–115936.748.333.3
54–96727–210420.040.031.7
104–1691430–381920.031.723.3
All201–381926.744.632.1

MolGoT is 11.6 to 28.3 pp above CoT in every group, and 17.9 pp above it overall (26.7% → 44.6%). Grouping the same molecules by molecular weight gives the same result:

Molecular weight (g/mol)MoleculesSMILES-CoTMolGoT
< 5003028.355.0
500–10003631.945.8
1000–20003330.337.9
≥ 2000219.538.1

MolGoT is higher in every weight range, and the advantage holds for the heaviest molecules: above 2000 g/mol, CoT solves 9.5% of the tasks and MolGoT solves 38.1%.

(2) The region decomposition remains useful for large molecules. Treating the whole molecule as a single region (last column of the first table) lowers accuracy in every size group, from 44.6% to 32.1% overall and from 31.7% to 23.3% in the largest group. The decomposition therefore still helps when molecules are large. The number of regions is a configurable limit, not a restriction on molecular size.

(3) The similarity term remains discriminative for large molecules. For each molecule, we take the candidates that MolGoT generated and measure the range of their similarity to the original molecule (maximum minus minimum). The average range is 0.12, 0.13, 0.17, and 0.12 in the four size groups, from the smallest to the largest. The range does not shrink as molecules grow, so the similarity term distinguishes the candidates of large molecules as well as those of small ones.

We will add these experiments to the revised paper.

W2 / Q2. A more detailed analysis of Fig. 5(a) and an extended ablation.

We thank the reviewer for this careful reading of Fig. 5(a). We agree that this ablation does not fully disentangle the components: its variants cover only part of the graph operations, and No merge selection removes only the final selection of composed candidates, not the composition itself. Below we first explain the three observations on Fig. 5(a), and then report a new ablation that removes each graph operation completely, one at a time and in combination.

(1) A closer analysis of Fig. 5(a).

  • The count route has the largest impact. The count route combines the counts of the individual regions into one answer, i.e., it aggregates regional evidence. The new ablation confirms this: without this aggregation of regional evidence, Qwen3-8B drops from 59.0% to 34.0% on the counting and understanding tasks (see (3)).
  • Merge selection has a limited effect. This variant only decides whether a composed candidate may be returned as the final answer. A composed candidate is returned only if it is at least as good as the best single-branch candidate (Eq. 2), which happened in 6 of the 990 editing traces of Table 32, so the final answer seldom changes.
  • Removing repair seemed to help some models. The difference from the full framework is not significant in eight of the nine models (p ≥ 0.36, exact McNemar test). The exception is R1, and it is an effect of truncation: in the original configuration, 1,033 of its 1,100 questions had at least one truncated call. With a larger output budget, removing repair lowers the accuracy of R1 by 18.3 pp (see (2)).

(2) New ablation that isolates each operation. We removed each graph operation of MolGoT completely, one at a time:

  • No branching: one candidate is kept per step instead of several; for the counting tasks, the molecule is not split into regions.
  • No scoring: candidates are selected at random instead of by their score; the validity check is retained.
  • No aggregation: several parent states (the evidence of different regions, or compatible edits) are no longer combined into one.
  • No refinement: invalid edits are no longer repaired, composed candidates are no longer revised, and disagreeing counts are no longer adjudicated.
  • No repair: only the repair of invalid edits is removed.

In each condition all other operations stay active. One further condition removes aggregation and refinement together, to measure how the two interact.

The ablation uses 110 questions, 10 for each of the eleven tasks in Table 1 of the paper, drawn from the same benchmark sources as the main experiments. The models are Qwen3-8B and R1; R1 uses a larger output budget than in the paper so that its reasoning is not truncated. All numbers are accuracy in %.

Branching, aggregation, and refinement apply to all 110 questions:

MolGoT variantQwen3-8BR1
Full38.650.9
No branching23.243.6
No aggregation27.742.7
No refinement37.340.0
No aggregation and no refinement28.234.5

Removing any of these operations lowers accuracy for both models. Qwen3-8B loses most from removing branching (15.4 pp) and aggregation (10.9 pp). R1 loses most from removing refinement (10.9 pp), and 16.4 pp when aggregation and refinement are both removed.

Scoring applies to 90 of the 110 questions, because the carbon- and ring-counting routes do not score candidates. On these 90 questions:

MolGoT variantQwen3-8BR1
Full28.945.6
No scoring20.016.7

Selecting candidates by score instead of at random is worth 8.9 pp for Qwen3-8B and 28.9 pp for R1.

Repair applies to the 60 editing and generation questions, because the other tasks do not execute edits. On these 60 questions:

MolGoT variantQwen3-8BR1
Full21.748.3
No repair19.230.0

Repair is worth 2.5 pp for Qwen3-8B and 18.3 pp for R1. The effect is smaller for Qwen3-8B because its repair step is rarely triggered.

W2 / Q2. A more detailed analysis of Fig. 5(a) and an extended ablation.

(3) Which part of the reasoning each operation affects, and how the operations interact.

  • Aggregation takes two forms. In the counting and understanding tasks, it combines the evidence of the individual regions into one answer; in the editing and generation tasks, it combines compatible edits of several candidates into one. Without aggregation, Qwen3-8B drops from 59.0% to 34.0% on the counting and understanding tasks, and R1 drops from 63.3% to 46.7% on the editing tasks.
  • Refinement and repair act on the editing and generation tasks. They revise a candidate using the feedback from executing the edit. Without refinement, R1 drops from 63.3% to 40.0% on editing and from 33.3% to 20.0% on generation, while its accuracy on the counting and understanding tasks stays at the same level (54.0% with refinement, 52.0% without).
  • Branching and scoring act on all tasks that search over candidates. Without either of them, accuracy drops on the counting and understanding, editing, and generation tasks alike.
  • Aggregation and refinement are complementary. For R1, removing aggregation costs 8.2 pp, removing refinement costs 10.9 pp, and removing both costs 16.4 pp, more than removing either one alone, so neither operation substitutes for the other.

We will add the new ablation and this analysis to the revised paper.

Reviewer_qBn2OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MolGoT uses molecular structure constraints to organize LLM reasoning and operations on candidate molecules. Taking SMILES as input, the system uses RDKit to establish persistent atom identifiers and structural regions, and selects reasoning paths according to the task. For editing and generation tasks, it uses executable JSON operations, candidate states, records of operations relative to the root structure, and mechanisms for combination and revision. Its main contribution is to turn the general Graphs of Thought approach into an executable, inspectable interface for molecular tasks, rather than to introduce an entirely new molecular representation. The paper evaluates the method across 11 task categories and 10 model configurations, and includes Molecule-of-Thought-style controls, comparisons using larger budgets, enhanced prompts, strong iterative baselines, and various ablations. In the main results table, MolGoT achieves the highest mean in six of the ten configurations and improves over the SMILES baseline in eight. Its advantages clearly depend on the model used.

Critical strengths
  1. The system interface is concrete, and its execution semantics are technically substantive.

    Pages 4–6 and Appendices A–B provide relatively detailed descriptions of persistent IDs, operation records relative to the root structure, namespaces for new IDs, candidate replay, and compatibility checks. The admission criterion in Equation (2) (p. 6) requires validity and no regression on the relevant metrics relative to the parent candidate. This does not guarantee correctness, but the process is easier to audit than free-text combination without execution checks.

  2. The empirical support goes beyond comparisons with simple prompting.

    In Table 2 (p. 7), Qwen3-8B improves from 20.73 with SMILES to 42.45 with MolGoT, while the GPT configuration improves from 74.45 to 80.64. The appendix further examines additional budgets and enhanced prompts on different subsets used for control comparisons. In Table 23 (p. 31), Qwen3-8B’s difference relative to the enhanced baseline with twice the budget is +16.73 percentage points, with an exploratory 95% interval of [13.09, 20.36]. These controls support the conclusion that the gains are not solely attributable to using more tokens.

  3. Some advantages persist after excluding specialized counting tasks.

    After excluding the carbon/ring tasks, Table 27 (p. 33) shows that Qwen3-8B, Qwen3-14B, and the Llama configuration still improve by 4.67, 5.33, and 7.56 percentage points, respectively, over a strong iterative baseline with shared guidance. The corresponding intervals all lie entirely above zero. This provides more targeted evidence of the system’s value for some models.

  4. The appendix clearly presents unfavorable results, helping to define the limits of the conclusions.

    In the same table, the R1 and DeepSeek configurations decline by 20.00 and 8.00 percentage points, respectively, while the interval for GPT’s difference includes zero. The paper also provides details on paired samples, bootstrap analyses, repeated runs, and failure cases. Overall, readers have the information needed to assess when the method works, without having to rely solely on the averages in the main table.

Critical weaknesses

1. Evidence for system-level gains is stronger than evidence that multi-parent combination itself is effective; attribution to this mechanism should therefore be more limited. — Insufficient evidence, without negating the overall contribution. In Table 28 (p. 34), removing merge changes performance by −0.18 percentage points for Qwen3-8B and +0.36 for GPT; removing repair actually improves R1 by 6.82 percentage points. In the descriptive analysis of the same trajectories in Appendix E.5 (p. 35), the default editing setting corrects none of the 594 incorrect single-candidate comparison cases, while worsening 4 of the 396 correct cases. The default generation setting also produces only a small number of corrections and regressions.

2. Online screening metrics do not fully align with the final task constraints, limiting the interpretation of “reliable control.” — A confirmed limitation in the scope of the implementation. Pages 5 and 19–20 explain that the online checks for some editing tasks accept any strict increase or decrease, whereas the final evaluation requires an exact change of ±1 together with corresponding structural evidence. For some generation tasks, the online checks use a lower bound on functional group counts, while the final evaluation requires exact counts. The SMARTS definitions and some structural checks also differ between online screening and final evaluation.

Questions for authors

Based on the existing comparisons with strong iterative baselines, can the authors clarify the conditions under which adopting MolGoT is appropriate?

6 comments

W1. Evidence that multi-parent combination itself is effective.

We thank the reviewer for this careful reading. Table 28 and Appendix E.5 do not fully show the contribution of multi-parent combination, because neither compares the framework with and without combination. We ran this comparison, and the result shows that multi-parent combination itself is effective.

New experiment. Multi-parent combination merges the edits of several existing candidates (the parents) into one new candidate. We stop the framework from creating such candidates, keep everything else unchanged, and compare the result with the full framework. The experiment uses 110 questions (10 for each of the eleven tasks in Table 1 of the paper, from the same benchmark sources as the main experiments). Multi-parent combination is used only in the editing and generation tasks, which account for 30 questions each. R1 uses a larger output budget than in the paper so that its reasoning is not truncated.

Result 1: with multi-parent combination, the editing accuracy of R1 is 16.7 pp higher. This experiment compares the full framework with the framework without multi-parent combination on the 30 editing questions (accuracy, %). We also split the questions by whether the full framework created a multi-parent candidate on them, 15 questions each.

R1, editing questionsFullNo multi-parent combination
All63.346.7
Combination used73.346.7
Combination not used53.346.7

The full framework is the best in all three rows. It is 26.7 pp higher on the questions where combination was used and only 6.7 pp higher on the other questions, so the gain comes from multi-parent combination itself.

Result 2: combined candidates are selected as final answers, and 83.3% of them are correct. This experiment records which kind of candidate becomes the final answer; for this purpose we also ran the 110 questions with GPT-5.6 Luna. On its 30 generation questions, 40.0% of the final answers are combined candidates, and 83.3% of them are correct. A combined candidate can also be closer to the goal than each of its parents: on one atom-number-constrained generation question of R1, the best individual candidate has an online goal score of 0.75 (the score measures how much of the goal a candidate satisfies), the combined candidate has 1.0, and the final answer is correct. Combined candidates are therefore used by the framework, and they are accurate.

Result 3: over all eleven tasks, accuracy is higher with aggregation (by 10.9 pp for Qwen3-8B and 8.2 pp for R1). Here we look at the same experiment over all 110 questions. Aggregation is multi-parent combination in the editing and generation tasks, and in the understanding tasks it combines the evidence from the individual regions of the molecule into one answer. Removing aggregation lowers accuracy from 38.6% to 27.7% for Qwen3-8B and from 50.9% to 42.7% for R1. Aggregation therefore improves accuracy across the whole benchmark.

Analysis of Table 28 and Appendix E.5. These three results in the paper do not show the effect of multi-parent combination, for the following reasons.

  • Table 28, No merge (−0.18 pp for Qwen3-8B, +0.36 pp for GPT). No merge does not remove multi-parent combination; it only excludes the combined candidates from the final selection, while combination runs as usual during the search (Appendix E.4). A result close to the full framework is therefore expected, and these two changes are not statistically significant (p ≥ 0.25, exact McNemar test on the same questions).
  • Table 28, No repair (+6.82 pp for R1). Repair is the step that fixes invalid edits, and it is a different component from multi-parent combination. This gain was obtained with truncated outputs and does not reflect the effect of repair: the output budget of R1 in the paper was too small, and 1,033 of its 1,100 questions had at least one call that was truncated before the answer was produced. With the larger output budget of the new experiment, removing repair lowers R1 from 48.3% to 30.0% on the editing and generation tasks, so repair helps.
  • Appendix E.5 (0 of 594 cases corrected). This analysis compares, within one run, the final answer with the best individual candidate, and counts a case as corrected when the former is correct and the latter is wrong; it does not rerun the framework without combination, as Appendix E.5 states. In the default setting, the final answer is an individual candidate in 99.4% of the runs on the editing tasks (Table 32), so the two answers being compared are the same and no case can be corrected. The analysis measures the final selection, whereas the new experiment measures the effect of combination.

W1. Evidence that multi-parent combination itself is effective.

In summary, the new experiment shows that multi-parent combination itself is effective: without it, the editing accuracy of R1 drops by 16.7 pp, and the loss is concentrated on the questions that use combination. We also identified why Table 28 and Appendix E.5 do not show this. We will add the new experiment and this analysis to the revised paper.

W2. Online screening metrics do not fully align with the final task constraints.

We thank the reviewer for pointing this out. The online checks and the final evaluation play different roles in the framework: the online checks guide the screening of candidates during the search, and the final evaluation decides whether the answer is correct after the search. For this reason, the online checks in the original implementation are looser than the final evaluation. All accuracies reported in the paper are computed by the final evaluation and are not affected by this difference. The online checks can also be made as strict as the final evaluation: an online check needs no reference answer and can verify a candidate by the standard of the final evaluation using only the input molecule and the goal stated in the question. We implemented such an exact check and reran the editing experiments: every output that passes the online check also passes the final evaluation, and accuracy rises by 2.9 pp.

The exact check. The exact check uses the definitions and procedures of the final evaluation directly, so the differences in SMARTS definitions and structural checks are removed as well. It requires that the count of the target group changes by exactly ±1; that there is structural evidence of the edit, that is, apart from the group that is added, deleted, or substituted, the output molecule is identical to the input molecule; and that groups are counted with the dictionary and counting function of the final evaluation. The final dictionary contains all 19 definitions of the online dictionary; it only changes the pattern of nitro and adds anhydride, sulfide, and borane (Appendix B.2). The exact check reads only the input molecule and the instruction.

Difference 1, editing tasks: any increase or decrease online, exactly ±1 with structural evidence in the final evaluation. This experiment compares the original check with the exact check: 120 editing questions (40 each for addition, deletion, and substitution; molecules with 9 to 169 carbon atoms), Qwen3-8B.

Online checkAccuracy (%)Share of outputs passing the online check that fail the final evaluation (%)
Original (paper)44.621.3
Exact47.50.0

The exact check is better. Accuracy rises by 2.9 pp, and it rises in each of the three tasks; the share of outputs that pass the online check but fail the final evaluation drops from 21.3% to 0, so the outputs that pass the online check are exactly the correct final outputs.

The table below splits the same outputs by the number of carbon atoms, to see whether the gain depends on molecule size (accuracy, %).

Carbon atomsOriginal checkExact check
9–2758.358.3
28–5248.351.7
54–9640.040.0
104–16931.740.0

The exact check is at least as accurate as the original check in every size range, and 8.3 pp higher for the largest molecules.

Difference 2, generation tasks: a lower bound online, exact counts in the final evaluation. This experiment changes the online check from "at least" to "exactly" and re-evaluates the outputs saved from our earlier runs of the functional-group-guided generation task (10 questions; Qwen3-8B, R1, and GPT-5.6 Luna).

Online checkShare of outputs passing the online check that fail the final evaluation (%)
Original (lower bound)27.3
Exact count0.0

With the exact count, the online check agrees completely with the final evaluation: the share of outputs that pass the online check but fail the final evaluation drops from 27.3% to 0. The lower bound is therefore the only cause of this difference.

In summary, the exact check resolves the mismatch between the online checks and the final evaluation: every output that passes the online check also passes the final evaluation, and accuracy rises by 2.9 pp. We will add the exact checks for the editing and generation tasks and these results to the revised paper, and describe the scope of "reliable control" accordingly.

Reviewer_LT9mOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolGoT, a training-free framework that organizes LLM-based molecular reasoning through structure-guided graphs of thought. It retains SMILES as the molecular interface and combines shared structural context, task-specific reasoning routes, and executable molecular edits with tool-based checking and conditional candidate composition and revision. Evaluations across various tasks and models show improvements over the baselines in several settings, although the benefits vary across models and tasks.

Critical strengths
  • Practical framework. MolGoT combines persistent atom identities, executable molecular edits, and tool-based checks within a training-free framework. Its recorded intermediate states make candidate generation, revision, and selection easier to inspect.
Critical weaknesses
  • Limited analysis of when the proposed method is effective. Table 2 shows that the effectiveness of MolGoT varies substantially across models. While it achieves the highest accuracy for Qwen3-8B and GPT, it underperforms MoleCode-GoT for the two smallest Qwen3 models, R1, and Gemini. Although the paper acknowledges this variation and examines some execution settings, it provides limited insight into why these differences arise and which model characteristics favor the proposed approach. A more systematic analysis would help clarify when MolGoT should be preferred over alternative reasoning strategies.
  • The contribution of each component is insufficiently isolated. For example, Figure 5(a) shows that the No merge selection and No repair variants achieve higher accuracy than the full framework for some model configurations, including R1 and GPT. These results suggest that these components do not provide consistent benefits across configurations. Clearer descriptions of what each ablation removes and retains, together with further analysis, would help establish the distinct contribution of each component and the conditions under which it is beneficial.
  • Insufficient justification for the scoring function in Eq. (1). The scoring function combines validity, constraint satisfaction, structural similarity, and locality using fixed weights of 0.35, 0.35, 0.15, and 0.15. However, the paper does not explain how these weights were chosen or report sensitivity analyses or ablations of the individual score terms. Given the score’s role in candidate ranking and merge acceptance, further analysis would help establish whether the resulting decisions and editing performance are robust to these design choices.
Questions for authors

Please address the concerns above.

5 comments

W1. Limited analysis of when the proposed method is effective.

We thank the reviewer for this comment and agree that the original analysis was limited. We first clarify the relation between the two GoT variants, and then give a systematic analysis.

(1) Both GoT variants are instances of our framework. MolGoT and Molecode-GoT apply our GoT framework to SMILES and MoleCode input, respectively; Molecode-GoT is not an external baseline. SMILES is a compact linear string in which the molecular topology is implicit; MoleCode [1] is a graph-explicit language that writes atoms and bonds as typed entities with explicit relations, so the topology can be read directly. Cases where Molecode-GoT exceeds MolGoT therefore show which input suits a model, not a failure of the framework. The table below reports the gain of GoT over CoT under the same input.

ModelSMILES: GoT − CoT (pp)MoleCode: GoT − CoT (pp)Average (pp)
Qwen3-0.6B−2.00+2.36+0.18
Qwen3-1.7B−1.64+7.45+2.91
Qwen3-4B+15.18+0.54+7.86
Qwen3-8B+21.720.00+10.86
Qwen3-14B+13.82+3.27+8.55
Llama-3.1-8B+6.91+2.73+4.82
R1+12.37+7.36+9.87
DeepSeek V4.1 Flash+2.64−1.18+0.73
Gemini 3 Flash+11.54+1.27+6.41
GPT-5.6 Luna+6.19+4.46+5.33

Averaged over the two inputs, GoT improves over CoT for all ten models.

(2) Which GoT variant performs better depends on the model's scale and on whether it performs extended reasoning on its own.

  • Small non-thinking models (Qwen3-0.6B, 1.7B). These models have limited capability and need the explicit structure: they benefit from GoT on MoleCode (+2.36 and +7.45 pp), where Molecode-GoT is the best of the four pipelines.
  • Non-thinking or low-reasoning models of 4B and above (Qwen3-4B, 8B, 14B, Llama-3.1-8B, DeepSeek with thinking disabled, GPT with low reasoning effort). These models do not reason at length on their own, but are capable enough to use the structural context and organized reasoning that MolGoT supplies. In all six, MolGoT is the best of the four pipelines (+2.64 to +21.72 pp over SMILES-CoT) and consumes fewer tokens than Molecode-GoT.
  • Thinking models (Gemini, R1). MoleCode is already much more accurate than SMILES under CoT (+14.45 pp for Gemini, +9.37 pp for R1): with SMILES, a thinking model spends its reasoning on reconstructing the molecular graph, whereas with MoleCode it can spend that reasoning on the task [1]. The same holds under GoT: Molecode-GoT is their best pipeline.

(3) Which GoT variant performs better also depends on the task. Averaged over models, Molecode-GoT is stronger on molecular formula, molecular equivalence, and carbon counting, where the explicitly written atoms help. MolGoT is stronger on ring counting, functional-group counting, and the three editing tasks, at a lower token cost.

Summary: when to prefer MolGoT.

  • By model: MolGoT for non-thinking or low-reasoning models of 4B and above (six of the ten models), where it gives the highest accuracy with fewer tokens; Molecode-GoT for small models and thinking models.
  • By task: MolGoT for ring counting, functional-group counting, and editing; Molecode-GoT for formula, equivalence, and carbon counting.

We will add the table and this analysis to the revised paper.

References

[1] Zhiyuan Yan, Chen Liu, Boxuan Zhao, Kaiqing Lin, Jixiang Zhao, Yimi Wang, Liuzhenghao Lv, Hao Li, Shanzhuo Zhang, Li Yuan, and Fanyang Mo. MoleCode unlocks structural intelligence in large language models. arXiv preprint arXiv:2605.16480, 2026.

W2. The contribution of each component is insufficiently isolated.

We thank the reviewer for pointing out this problem in the ablation of Fig. 5(a). We analyzed that ablation in more detail and found that the higher accuracy of the No merge selection and No repair variants comes from random variation and, in one case, from truncated outputs. We also agree that the ablation does not fully isolate the components, and we have added a new ablation.

(1) A closer analysis of Fig. 5(a). We compared the No merge selection and No repair variants with the full framework for GPT and R1, the two models named by the reviewer, on the same questions (exact McNemar test):

VariantModelDifference from fullp
No merge selectionGPT+0.36 pp0.87
No merge selectionR1+1.82 pp0.61
No repairGPT+0.73 pp0.61

None of these differences is significant, and the same holds for the other seven models (p ≥ 0.25). The remaining case, No repair for R1 (+6.82 pp), is an effect of truncation: in the original R1 configuration, 1,033 of 1,100 questions had at least one truncated call. In the new ablation of (2), which uses a larger output budget, removing repair lowers the accuracy of R1 by 18.3 pp, so repair is beneficial for R1 once its outputs are complete.

(2) New ablation that isolates each operation. The variants in Fig. 5(a) do not fully isolate the components: No merge selection removes only the final selection of composed candidates, not the composition itself, and the variants cover only part of the graph operations. We therefore removed each graph operation of MolGoT completely, one at a time:

  • No branching: one candidate is kept per step instead of several; for the counting tasks, the molecule is not split into regions.
  • No scoring: candidates are selected at random instead of by their score; the validity check is retained.
  • No aggregation: several parent states (the evidence of different regions, or compatible edits) are no longer combined into one.
  • No refinement: invalid edits are no longer repaired, composed candidates are no longer revised, and disagreeing counts are no longer adjudicated.
  • No repair: only the repair of invalid edits is removed.

In each condition all other operations stay active. One further condition removes aggregation and refinement together.

The ablation uses 110 questions, 10 for each of the eleven tasks in Table 1 of the paper, drawn from the same benchmark sources as the main experiments, with Qwen3-8B and R1. R1 uses a larger output budget than in the paper so that its reasoning is not truncated.

Branching, aggregation, and refinement apply to all 110 questions (accuracy in %):

MolGoT variantQwen3-8BR1
Full38.650.9
No branching23.243.6
No aggregation27.742.7
No refinement37.340.0
No aggregation and no refinement28.234.5

Removing any of these operations lowers accuracy for both models. For Qwen3-8B, the largest drops come from removing branching (38.6% to 23.2%) and aggregation (to 27.7%). For R1, the largest drop comes from removing refinement (50.9% to 40.0%), and removing aggregation and refinement together lowers accuracy further, to 34.5%.

Scoring applies to 90 of the 110 questions; the carbon- and ring-counting routes do not score candidates (accuracy in % on these 90 questions):

MolGoT variantQwen3-8BR1
Full28.945.6
No scoring20.016.7

Without scoring, accuracy drops from 28.9% to 20.0% for Qwen3-8B and from 45.6% to 16.7% for R1.

Repair applies to the 60 editing and generation questions; the other tasks do not execute edits (accuracy in % on these 60 questions):

MolGoT variantQwen3-8BR1
Full21.748.3
No repair19.230.0

Without repair, accuracy drops from 48.3% to 30.0% for R1 and from 21.7% to 19.2% for Qwen3-8B.

(3) When each operation is beneficial.

  • Aggregation takes two forms. In the counting and understanding tasks, it combines the evidence of the individual regions into one answer, which is the role of the counting routes in Fig. 5(a); No merge selection does not remove this step. In the editing and generation tasks, it combines compatible edits of several candidates into one. Without aggregation, Qwen3-8B drops from 59.0% to 34.0% on the counting and understanding tasks, and R1 drops from 63.3% to 46.7% on the editing tasks.
  • Refinement and repair act on the editing and generation tasks. They revise a candidate using the feedback from executing the edit. Without refinement, R1 drops from 63.3% to 40.0% on editing and from 33.3% to 20.0% on generation.
  • Branching and scoring act on all tasks that search over candidates. Without either of them, accuracy drops on the counting and understanding, editing, and generation tasks alike.

We will add it in revision.

W3. Insufficient justification for the scoring function in Eq. (1).

We thank the reviewer for this comment. Below we explain how the weights were chosen, and we add a sensitivity analysis of the weights and an ablation of the similarity term. Accuracy and merge acceptance are robust to the weights.

(1) How the weights were chosen. The four terms have two levels of priority. Validity V and the visible task predicate C_vis are the primary criteria: whether a candidate is a valid molecule and whether it makes progress on the public constraints. Similarity S and locality L are secondary: they prefer candidates that preserve the rest of the molecule and confine the edit to the target regions. The weights were chosen so that this priority holds in the two places where the score is used:

  • Candidate ranking. As stated in Section 3.4, candidates are compared on validity and visible progress before the composite score. S and L therefore only decide between candidates that are equal on validity and progress.
  • Merge acceptance. A composed candidate is accepted only if its progress and its composite score are not lower than those of the best individual candidate (Eq. 2). Each primary criterion has a larger weight (0.35) than S and L together (0.15 + 0.15), so a candidate that satisfies the public constraints always scores higher than one that satisfies none of them, even if its similarity and locality are lower.

(2) Sensitivity to the weights. We ran MolGoT with four weight settings on 300 editing questions (100 each of addition, deletion, and substitution) with Qwen3-8B:

Weights (V / C_vis / S / L)Accuracy (%)
0.35 / 0.35 / 0.15 / 0.15 (paper)35.1
0.40 / 0.40 / 0.10 / 0.1035.1
0.35 / 0.35 / 0.20 / 0.1035.2
0.35 / 0.35 / 0.10 / 0.2035.3

Accuracy changes by at most 0.2 pp, and no setting differs significantly from the paper's weights (p ≥ 0.80, exact McNemar test). Merge acceptance does not change either: we re-evaluated the 4,190 merge and repair proposals recorded in these runs under all four settings, and the same 200 proposals are accepted under each. All four settings keep all four terms and keep the weight of each primary criterion above the sum of the weights of S and L. Under these two conditions, the editing accuracy and the merge decisions are robust to the exact values of the weights.

(3) Ablation of the similarity term. We set the weight of the similarity term to zero and kept the other three weights, on 90 editing questions (30 each of addition, deletion, and substitution) with two models (accuracy in %):

ModelPaper weightsNo similarity
Qwen3-8B41.726.7
Qwen3-4B28.916.1

Without the similarity term, accuracy drops from 41.7% to 26.7% for Qwen3-8B and from 28.9% to 16.1% for Qwen3-4B (p < 0.001 for each, exact McNemar test).

We will add the explanation of the weights, the sensitivity analysis, and the ablation to the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.