acceptodds
Under review as a conference paper at ICLR 2027

Suggest-Review-Edit-Verify: Structure-Preserving Edit-Space Reduction for Iterative Refinement in Agentic Molecular Generation

11 views

Abstract

Molecule generation, which aims to discover molecules with desired characteristics, has recently advanced through natural-language instructions that specify diverse design objectives. Recent LLM-based chemistry agents make generation actionable by leveraging these instructions for iterative generation and refinement. Nonetheless, existing methods struggle to identify effective modifications, limiting their efficacy for jointly satisfying structural requirements and multiple property constraints. To address this challenge, we introduce SREV-Mol (Suggest–Review–Edit–Verify) for Molecular Generation), a structure-preserving framework for iterative molecular refinement. Our approach first translates the natural-language structural requirement into SMARTS (SMILES arbitrary target specification) and generates a verified seed molecule, then iteratively selects effective edits from a reduced space of structure-preserving candidates to refine the molecule. To further improve edit selection, we present Structured Edit Reasoning (SER) and candidate-specific Auxiliary Feature Tokens (AFTs), which provide complementary signals for comparing candidate modifications. Thereby, SREV-Mol enables effective multi-property optimization under structural requirements and up to 14 simultaneous property constraints, substantially reducing broad trial-and-error search over the molecule edit space. Extensive experiments on in-distribution (ID) and scaffold out-of-distribution (OOD) settings show that ours outperforms existing chemistry agents and strong frontier LLMs, despite using an 8B backbone, while continuing to improve with larger refinement budgets.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jnxfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SREV-Mol, a molecular generation framework designed to satisfy natural-language structural requirements and multiple property constraints. The framework first translates structural descriptions into SMARTS and generates verified seed molecules. It then uses matched molecular pair (MMP) statistics to construct a structure-preserving edit space, while Structured Edit Reasoning (SER) and candidate-aligned Auxiliary Feature Tokens (AFTs) assist the language model in selecting molecular modifications through iterative verification.

Experiments on 900 in-distribution (ID) and 300 scaffold out-of-distribution (OOD) tasks report joint success rates of 67.8% and 29.0%, respectively, outperforming the strongest reported baselines at 59.1% and 24.0% in the corresponding settings.

Critical strengths

S1. Meaningful problem setting and competitive end-to-end performance.

The paper addresses a practically relevant molecular constraint-satisfaction problem by combining natural-language structural requirements with multiple property intervals. Tables 2 and 3 demonstrate that SREV-Mol achieves the highest joint success rates among the evaluated methods on both ID and Scaffold-OOD benchmarks. These results provide positive evidence for the effectiveness of the complete system on the proposed tasks.

S2. Empirical evidence supporting MMP-based search-space reduction and structural preservation.

Figure 3(a) shows that random editing within the MMP candidate space achieves 64.7% success when reference SMARTS and ground-truth seeds are provided, substantially outperforming the evaluated generic random-editing strategies. Table 4 further demonstrates that removing structure-preserving filtering reduces ID joint success from 67.8% to 47.6%. These results provide convincing evidence that incorporating structural constraints and property-informed edit statistics is beneficial for molecular refinement.

Critical weaknesses

W1. The independent contribution and necessity of SER/AFT-based agentic review are not sufficiently established through end-to-end evaluations.

SER and AFTs are presented as central technical contributions. However, the ablations in Section 4.3.2 and Figure 3(b) are conducted only on a separate Qwen3-1.7B model and evaluate step-level edit-selection accuracy rather than the final success of the complete Qwen3-8B system. Although the feature perturbation experiments in Figure 11(c) indicate that the model uses candidate-specific AFT information, they do not establish how much SER and AFTs independently improve end-to-end generation success.

This concern is particularly important because Table 2 shows that, compared with Claude-Opus-5.0 + SREV, SREV-Mol improves structural satisfaction from 70.8% to 82.8%, whereas complete property satisfaction increases only from 79.9% to 80.6%. Consequently, the overall improvement may be primarily driven by the specially trained structural interpretation and seed-generation stage rather than more effective edit selection.

Furthermore, Appendix C.4 already introduces a pretrained candidate-value predictor for estimating downstream editing success, yet the paper does not evaluate a non-LLM policy that directly ranks candidates using this predictor. Figure 3(a) reports 94.5% success for MMP-based beam search. Although this controlled experiment uses reference SMARTS and ground-truth seeds and is therefore not directly comparable to the end-to-end setting, it suggests that the underlying MMP search space already provides substantial optimization capability.

The paper therefore does not adequately answer a fundamental question: What additional benefit does LLM-based Review provide beyond simpler MMP search or learned value-based candidate selection?

W2. Baseline coverage and evaluation design are insufficient to fully substantiate the claimed methodological advantages and generalization.

First, although the related-work section discusses OptADMET, M4olGen, MolReAct, and PROBE, the main experiments do not directly compare against these closely related editing or optimization approaches under matched seeds, property evaluators, and verification budgets. Strong non-agent optimization baselines are also missing. Moreover, the evaluated chemistry agents exclude some inaccessible tools from their original implementations. Consequently, their near-zero joint success rates do not, by themselves, establish superiority over related molecular optimization methods. Even if these methods cannot directly handle the complete natural-language task, controlled optimization experiments starting from common seeds would provide a more meaningful comparison.

Second, Appendix E.3 explicitly states that the edit-selection dataset is randomly split at the individual-round level rather than by source task or complete trajectory. Since each example retains its full preceding trajectory context, highly correlated states originating from the same task may appear in both training and test sets. This does not demonstrate actual data leakage, but it weakens the independence of the evaluation used to support SER/AFT generalization.

Finally, property constraints are constructed around feasible reference molecules, and test structural descriptions undergo selection, with approximately 25–35% of candidate descriptions passing round-trip verification. While this construction is reasonable for ensuring benchmark consistency, it limits the conclusions that can be drawn about unfiltered natural-language requests. Moreover, despite emphasizing support for up to 14 simultaneous property constraints, the paper does not report success rates stratified by constraint count. Paired uncertainty estimates or statistical significance tests for the main OOD differences are also absent.

These limitations do not invalidate the reported benchmark results, but they substantially restrict the strength of the broader methodological and generalization claims.

Questions for authors

Q1. Can the authors provide full Qwen3-8B end-to-end ablations comparing Direct Review, SER-only, SER+AFT, direct AFE value-predictor ranking, and MMP greedy selection under matched Stage 1, training data, initial seeds, and verification budgets? Such experiments are essential for identifying the independent contribution of the proposed review components.

Q2. Can the edit-selection evaluation in Figure 3(b) be repeated using task-disjoint or trajectory-disjoint train/test splits to prevent correlated states from the same trajectories from appearing across splits? Do the gains from SER and AFTs remain consistent?

Q3. Can the authors report performance by the number of simultaneous property constraints, particularly for 10–14 constraints, provide paired statistical analyses on Scaffold-OOD, and clarify the absence of task-aligned non-agent or related fragment-editing baselines?

7 comments

We sincerely thank the reviewer for the detailed and constructive feedback and valuable suggestions for strengthening our component-level validation, baseline comparisons, and generalization analysis.

1. Independent Contributions of SER and AFTs Beyond Stage 1 (W1, Q1)

What additional benefit does LLM-based Review provide beyond simpler MMP search or learned value-based candidate selection? … Can the authors provide full Qwen3-8B end-to-end ablations comparing Direct Review, SER-only, SER+AFT, direct AFE value-predictor ranking, and MMP greedy selection under matched Stage 1, training data, initial seeds, and verification budgets?

→ To isolate the contribution of Review to property optimization, we conducted an additional controlled evaluation on all 1,200 instances (900 ID and 300 Scaffold-OOD). Specifically, we first trained each method end-to-end using Qwen3-8B under identical training settings, and matched Stage 1 to ensure that all methods started from the same initial seed molecules and compared their optimization performance. We then evaluated only the Stage 2 optimization loop with a maximum of six edits, using the same Suggest candidate space and verification procedure.

We compared five methods: MMP Greedy, which selects the top-ranked Suggest candidate w/o Review; direct AFE value-predictor ranking, which selects the candidate with the highest pretrained success score; Direct Review, which fine-tunes the LLM to select edits without SER or AFTs; SER-only, which incorporates structured reasoning without AFTs; and SER+AFT, which represents the full SREV-Mol model. All LLM-based variants used the same Qwen3-8B backbone, training data, and training settings, varying only the Review components.

Review TerminologyPaper TerminologySucc.↑(=Prop.↑)# Edits ↓# Verify ↓
MMP GreedySuggest Greedy61.92 (72.11 / 31.33)3.76 (3.38 / 4.91)4.76 (4.38 / 5.91)
direct AFE value-predictor rankingCandidate-value prediction model70.42 (80.11 / 41.33)3.45 (3.08 / 4.57)4.45 (4.08 / 5.57)
Direct ReviewSREV-Mol (w/o SER, w/o AFT)69.75 (79.33 / 41.00)3.51 (3.12 / 4.69)4.51 (4.12 / 5.69)
SER-onlySREV-Mol (w/o AFT)70.50 (81.00 / 39.00)3.52 (3.05 / 4.66)4.52 (4.05 / 5.66)
SER+AFTSREV-Mol72.17 (82.78 / 40.33)3.41 (2.99 / 4.65)4.41 (3.99 / 5.65)

Table R14. Stage 2 optimization performance under matched structural inputs and verification budgets. The values separated by "/" within the parentheses represent the ID and Scaffold-OOD scores, respectively.

With Stage 1 fixed, SREV-Mol outperforms MMP Greedy by 10.25 percentage points (72.17% vs. 61.92%), demonstrating the benefit of learned candidate selection (Table R14). Direct AFE value-predictor ranking also achieves strong performance (70.42%), highlighting the utility of pretrained candidate-value prediction. The full model further improves success by 1.75 percentage points overall, with incremental gains from Direct Review (69.75%) to SER-only (70.50%) and SER+AFT (72.17%). Although AFE ranking performs slightly better on Scaffold-OOD (41.33% vs. 40.33%), SREV-Mol shows a larger advantage on tasks with 10–14 simultaneous property constraints (70.8% vs. 65.0%, +5.8 points; Table R20, Response 6). These results suggest that SER and AFTs provide additional benefits for challenging multi-property optimization beyond Stage 1 structural generation.

We will include these ablation results in the revised paper.

2. Task-Aligned Optimization Baselines (W2, Q3)

First, although the related-work section discusses OptADMET, M4olGen, MolReAct, and PROBE, the main experiments do not directly compare against these closely related editing or optimization approaches under matched seeds, property evaluators, and verification budgets. Strong non-agent optimization baselines are also missing. … and clarify the absence of task-aligned non-agent or related fragment-editing baselines?

→ We first clarify that our task aims to jointly satisfy complex natural-language structural requirements and multiple property constraints, rather than optimizing molecular properties alone.

Although OptADMET provides public code, its required databases, model weights, and some implementation files are unavailable. Similarly, executable implementations or checkpoints for M4olGen and MolReAct were unavailable. We therefore additionally evaluated PROBE, structure-based drug design agent, by initializing its InteractionAgent with ground-truth seed SMILES, replacing its Vina-based objective with our normalized property-gap objective, and excluding edits that violate the reference SMARTS. Since each PROBE loop involves approximately ten editing and verification operations, we allowed up to six loops rather than six individual edits.

MethodSucc.↑Struct.↑STRatio↑Prop.↑PTRatio↑Gain↑
PROBE17.097.997.917.075.735.2
SREV-Mol67.882.885.280.696.690.5

Table R15. Comparison with task-adapted PROBE on 900 ID instances.

PROBE achieves lower property satisfaction despite even starting from ground-truth seeds and using additional internal edits(Table R15).

We additionally evaluated GraphGA, REINVENT, and GPBO, the top three methods in the PMO benchmark. We adapted them to start from ground-truth seed SMILES while preserving the required structure. REINVENT used seed-conditioned generation with varying attachment sites, while GraphGA and GPBO restricted modifications to non-seed regions using ZINC-derived BRICS fragments. All methods used our normalized property-gap objective and evaluators. We measured the evaluation budget by the number of times the property evaluators were called to assess candidate molecules during optimization.

# Property EvaluationsGraphGAREINVENTGPBOSREV-Mol
66.211.88.667.8
1010.016.311.2-
10037.645.136.2-
50066.867.173.7-
1,00078.977.880.2-
5,00095.888.094.0-
10,00098.288.296.4-
50,00098.988.296.4-

Table R16. Table R16. Joint success rates (Succ.) under varying property evaluation budgets during optimization on 900 ID instances. Structural satisfaction is enforced at 100%.

Although we adapted these baselines to compare their optimization capabilities under as closely matched conditions as possible, a fully fair comparison remains challenging due to differences in their optimization mechanisms and computational costs. In particular, the reported evaluation budgets only account for property evaluations during optimization, excluding those used for SREV-Mol's MMP library construction and training-data generation. We will clarify these limitations in the revised paper.

3. Clarification of Excluded Chemistry Agent Tools

Moreover, the evaluated chemistry agents exclude some inaccessible tools from their original implementations.

Regarding the excluded chemistry-agent tools, we omitted RXN4Chem (forward synthesis and retrosynthesis), ChemSpace (reactant-price lookup), and Tavily (web search) due to API accessibility or usage limitations. Specifically, RXN4Chem and ChemSpace were not accessible through available API credentials, while Tavily's limited API quota was insufficient for our evaluation. These tools are primarily intended for synthesis-related tasks or external information retrieval, rather than direct molecular generation and property optimization.

4. Train-Test Independence in Edit-Selection Evaluation (W2, Q2)

Can the edit-selection evaluation in Figure 3(b) be repeated using task-disjoint or trajectory-disjoint train/test splits?

→ To address potential correlations in the original round-level split, we repeated Figure 3(b) using trajectory-disjoint training and test sets, sampling 100K and 5K rounds, respectively. All other settings remained unchanged, and results were averaged over four random seeds.

MethodOracle Acc. (%)Reference Acc. (%)Oracle Acc. Hard (%)Reference Acc. Hard (%)
Direct (SREV-Mol* (w/o SER, w/o AFT))48.2 ± 0.835.7 ± 2.745.8 ± 1.435.6 ± 2.0
Freeform CoT44.0 ± 1.230.5 ± 0.541.8 ± 4.530.2 ± 5.1
Freeform CoT (w/ features)42.1 ± 1.330.9 ± 0.435.8 ± 3.023.9 ± 4.1
SREV-Mol* (w/o AFT)51.7 ± 1.140.9 ± 0.649.4 ± 1.240.1 ± 0.9
SREV-Mol*52.6 ± 1.341.6 ± 1.051.8 ± 1.742.5 ± 2.1

Table R17. Edit-selection accuracy under trajectory-disjoint train/test splits.

The performance trends remain consistent with Figure 3(b): SER-only outperforms Direct and Freeform CoT, while AFTs further improve all four accuracy metrics(Table R17). These results confirm that the benefits persist under trajectory-disjoint evaluation.

5. Effect of Round-Trip Verification Filtering (W2)

Test structural descriptions undergo selection, with approximately 25–35% of candidate descriptions passing round-trip verification. … it limits the conclusions that can be drawn about unfiltered natural-language requests.

→ We introduced round-trip verification to ensure that natural-language structural descriptions faithfully correspond to their reference SMARTS, enabling reliable evaluation of structural satisfaction. To examine the effect of this filtering process, we constructed two sets of 1,000 instances each: RTPass, containing descriptions that passed round-trip verification, and RTFree, containing descriptions generated without applying round-trip verification.

We first compared the structural distributions of the two datasets.

Structural FeatureDatasetMeanMax
# StereocentersRTPass1.2010
RTFree3.1726
# Heavy atomsRTPass33.171
RTFree38.173
Molecular weight (g/mol)RTPass466983
RTFree537999
# Rings (SSSR)RTPass4.1111
RTFree4.8230

Table R18. Structural distribution comparison between RTPass and RTFree.

RTFree contains more complex structures, particularly in stereocenters and ring systems (Table R18). However, RTPass still includes molecules with up to 71 heavy atoms, 10 stereocenters, and 11 rings, indicating that round-trip filtering does not restrict evaluation to simple molecules.

To further examine whether round-trip filtering affects the types of structural errors encountered by SREV-Mol, we evaluated Stage 1 (Translation and seed generation) on both datasets and categorized seed-generation failures relative to the reference structures.

Error TypeRTPassRTFree
Correct composition, incorrect arrangement260 (72.4%)391 (50.6%)
Failed to generate a molecule32 (8.9%)119 (15.4%)
Incorrect ring count32 (8.9%)112 (14.5%)
Incorrect ring composition (size/aromaticity/heteroatoms)33 (9.2%)98 (12.7%)
Other2 (0.6%)52 (6.8%)

Table R19. Distribution of Stage 1 structural failure types for RTPass (359 failures) and RTFree (772 failures). Percentages are calculated over failed instances.

Stage 1 achieves 64.1% structural success on RTPass, compared with 22.8% on RTFree. This gap may reflect both the greater structural complexity of RTFree and potential inconsistencies between unfiltered descriptions and their reference SMARTS. Nevertheless, incorrect structural arrangement remains the dominant failure type in both datasets(Table R19).

These findings highlight the importance of round-trip filtering for ensuring reliable structural evaluation, while also showing that it affects the difficulty and distribution of test instances. Therefore, our filtered benchmark provides a more controlled evaluation but does not fully represent unrestricted natural-language requests. We will clarify this trade-off and its limitations in the revised paper.

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

This paper presents SREV-Mol, a molecular optimization framework based on the Suggest–Review–Edit–Verify workflow, with structured reasoning and auxiliary features for edit selection. Experiments on ID and scaffold-OOD datasets show promising results compared with existing chemistry agents and frontier LLMs.

Critical strengths
  • The workflow is carefully designed, particularly the structure-preserving edit-space reduction and iterative verification.

  • The experimental results are promising, especially considering the relatively small 8B backbone.

Critical weaknesses
  • Limited generalization validation. The natural-language requirements are synthetically generated from molecular structures (i.e., Qwen3.6-27B). It is unclear how well the model handles real human-written requirements or how reliable and diverse the generated descriptions are.

  • Scaffold-disjoint OOD evaluation does not account for molecular similarity.

  • Insufficient analysis of reasoning and efficiency. The chemical correctness of the Review reasoning is not directly evaluated

  • The paper also lacks token-efficiency and inference-cost comparisons, making it difficult to assess the practical advantages over frontier LLMs.

Questions for authors
  • Have the authors evaluated the framework on human-written molecular design requirements? How is the quality of the synthetic descriptions assessed?

  • Could the authors report molecular similarity between training and OOD test molecules and analyze performance across similarity ranges?

  • Could the authors report actual reasoning-token consumption and inference costs? How sensitive are the results to reasoning budgets?

  • Are the chemical explanations generated during Review scientifically correct? Could the authors provide an expert evaluation or additional ablations?

9 comments

We sincerely thank the reviewer for the thoughtful feedback and valuable suggestions regarding generalization, reasoning reliability, and inference efficiency.

1. Synthetic Description Quality and Human-Written Generalization

Have the authors evaluated the framework on human-written molecular design requirements? How is the quality of the synthetic descriptions assessed?

→ To minimize errors in synthetic descriptions, we provided the description-generating LLM with accurate RDKit-derived structural information, including atom counts, functional groups, substitution positions, and linker connectivity, as part of the prompts during dataset construction.

Quality of synthetic descriptions. To assess the quality of descriptions generated by Qwen3.6-27B for training, we asked three external researchers with chemistry expertise to evaluate 60 description–SMARTS pairs (20 per evaluator). Each description was evaluated and categorized as "Accurate," "Partly Inaccurate," or "Wrong" based on whether it precisely represented the corresponding SMARTS structure.

RatingCountPercentage
Accurate5185
Partly Inaccurate915
Wrong00

Table R7. Expert evaluation of synthetic structural-description accuracy (60 instances).

Of the 60 descriptions, 51 (85%) were rated as accurate, while the remaining nine were partly inaccurate(Table R7). We further categorized the errors in these nine descriptions.

Error typeCount
Wrong ring or skeleton name (including isomers)0
Wrong aromaticity0
Wrong heteroatom type or count0
Wrong substitution or ring connection position in rings7
Wrong linker description2
Something in the substructure is missing from the description0
The description adds something that is not in the substructure2

Table R8. Error categories in partly inaccurate descriptions (multiple categories allowed per instance).

Most errors involved substitution or connection positions in rings(Table R8), suggesting that accurately expressing positional relationships in molecular graphs remains challenging, even with RDKit-derived structural information.

Description diversity and generalization to human-written requirements. To reduce dependence on a single description generator, we used different LLMs for training and evaluation descriptions (Qwen3.6-27B and DeepSeek-V4-Pro, respectively). We further examined generalization across description sources by asking the three chemistry researchers to independently write structural descriptions for the same 30 instances. We evaluated SREV-Mol using both the LLM(DeepSeek-V4-Pro)-generated and human-written descriptions while keeping the underlying molecular design tasks identical.

DescriptionSucc.↑Struct.↑STRatio↑Prop.↑PTRatio↑Gain↑# Edits↓
LLM-generated86.793.393.390.095.594.02.70
Human-written66.773.377.580.090.087.12.73

Table R9. SREV-Mol performance on 30 paired molecular design instances using LLM-generated and high-quality human-written structural descriptions.

These results indicate that SREV-Mol generalizes to human-written structural descriptions to some extent, achieving 66.7% joint success despite a performance drop compared with LLM-generated descriptions(Table R9). While property satisfaction remains relatively high (80.0%), the larger decrease in structural satisfaction (93.3% to 73.3%) suggests that interpreting diverse human-written structural requirements remains a key area for improvement. We will incorporate these findings and discuss the remaining limitations in the revised paper.

2. Scaffold Similarity and OOD Generalization

Could the authors report molecular similarity between training and OOD test molecules and analyze performance across similarity ranges?

→ To further assess generalization beyond scaffold disjointness, we measured scaffold overlap and scaffold-level Tanimoto similarity using Morgan fingerprints (ECFP4). For each test instance, we computed the mean similarity between its scaffold and all training scaffolds, then averaged these values across the Scaffold-OOD test set (Table R10).

DatasetScaffold overlap (%)Mean Tanimoto similarity
Scaffold-OOD0.00.1248

Table R10. Scaffold overlap and mean scaffold similarity with the training data.

We further evaluated joint success(Succ.) across four similarity intervals, dividing the 300 Scaffold-OOD instances into equally sized groups based on their mean similarity to the training scaffolds.

< 0.1120.112–0.1290.129–0.142≥ 0.142
# Instances75757575
Claude-Opus-5.0 (w/ SREV)22.713.318.724.0
GPT-5.6-Sol (w/ SREV)18.713.318.713.3
SREV-Mol22.737.324.032.0

Table R11. Joint success rates (Succ.) across mean scaffold-similarity intervals on Scaffold-OOD.

SREV-Mol achieves the highest or tied-highest success rate across all four similarity intervals(Table R11), outperforming both baselines in three intervals and matching Claude in the lowest-similarity group. Moreover, its performance does not exhibit a monotonic decline as mean scaffold similarity decreases. These results show that SREV-Mol's performance advantage over the baselines is maintained across the evaluated similarity ranges, although further evaluation is needed to establish generalization to more structurally distant molecules.

We will include these analysis in the revised paper.

3. Inference Cost and Reasoning-Budget Sensitivity

Could the authors report actual reasoning-token consumption and inference costs? How sensitive are the results to reasoning budgets?"

→ To examine inference efficiency and sensitivity to reasoning budgets, we evaluated GPT-5.6-Sol and Claude-Opus-5.0 under low, medium, and high reasoning-effort settings on SREV framework. We randomly sampled 100 ID instances and evaluated both models using the SREV framework.

For comparison, SREV-Mol was evaluated using vLLM on a single NVIDIA B200 GPU, processing one instance at a time. Since reasoning-token counts were not separately available for Claude, we consistently report total output tokens for all models.

Reasoning EffortModelOutput Tokens↓Solving Time (sec/instance ↓API Cost ($/instance)↓Succ.↑
LowGPT-5.6-Sol183532.80.08548.0
LowClaude-Opus-5.0189829.90.11953.0
MediumGPT-5.6-Sol280939.90.10451.0
MediumClaude-Opus-5.0344751.00.16755.0
HighGPT-5.6-Sol7167148.70.21848.0
HighClaude-Opus-5.09516130.60.40461.0
N/ASREV-Mol219813.9N/A (local)64.0

Table R12. Inference efficiency and success rates across reasoning-effort settings.

On the sampled 100 ID instances, SREV-Mol achieves 64% success with average 2,198 output tokens per instances and an average latency of 13.9 seconds per instance on a single B200 GPU(Table R12). In comparison, the frontier LLMs achieve 48-61% success while generally incurring higher token consumption and latency. Although inference times are measured under different serving environments and are not hardware-normalized, these results demonstrate favorable performance and latency for SREV-Mol in the evaluated setup.

Unlike frontier LLMs, since SREV-Mol’s Review format is fixed, SREV-Mol does not have an explicit reasoning-effort setting. Instead, its test-time computation can be scaled by increasing the number of Suggest-Review-Edit-Verify iterations. To further examine this behavior, we evaluated the same 100 ID instances under different maximum edit budgets. For GPT-5.6-Sol and Claude-Opus-5.0, we used medium reasoning effort to limit API costs.

Model / Max # Edits123456789101112
GPT-5.6-Sol333743495151525252525252
Claude-Opus-5.0354147485055575959595959
SREV-Mol213547565864646666666969

Table R13. Success rates (Succ.) under different maximum edit budgets.

Under matched edit budgets, SREV-Mol surpasses both frontier LLM baselines from four edits onward, reaching 64% success at six edits and 69% at eleven edits(Table R13). This improvement demonstrates that SREV-Mol can effectively utilize additional refinement steps(reasoning budget) to improve constraint satisfaction.

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

The authors propose a framework to generate molecules satisfying multiple desired properties from a natural language prompt. The framework starts by generating a molecular structure matching the natural language prompt, identifies the 4 best matching possible edits from a list precomputed from molecular data, then chooses among these edits based on their statistics and a learned representation predicting how often an edit leads to a successful molecule within a depth-5 search.

Critical strengths
  • The paper addresses a specific application domain with a clear idea (i.e., restraining the edit space of the agentic system in an informed manner).
  • The methodology is principled and well motivated: all the components of the method have clear motivations and specific goals, despite the complexity of the resulting method (see weaknesses).
Critical weaknesses
  • (W1) The paper does not show that the learned Review step beats a simple search over the same candidate set.: SREV-Mol was never run within the set up of Figure 3 (a) (i.e. with the SMARTS structure and the seed molecule provided), so it is unclear how much the learned components of SREV-Mol add compared to more basic edit selection methods (e.g. random, beam search, exhaustive etc). I suggest the authors add this ablation, because if the learned Review does not clearly beat greedy selection under matched conditions, the paper's central contribution is unsupported.

  • (W2) The review module is fairly complex; the paper is missing ablations to empirically justify the design.: one ablation I suggest is to include the score output you get during pretraining (Eq. 11–13) in the prompt, instead of using the learned embedding vector. If adding the scores works, it would be possible to supply it to other LLMs (i.e. closed models available only through an API), and would simplify the method by removing the projector, the placeholder tokens with embedding injection, and the joint training of the projector with the language model. To be clear, I am asking for the single pretrained score given to the fine-tuned 8B model in place of the AFTs, which is different from the ablation in Appendix F.2 (Tables 9-10, supplying textual features to off-the-shelf LLMs).

  • (W3) Most of the empirical gains are on structural satisfaction when the method is supposed to be optimizing for properties.: Table 2 shows that property satisfaction is almost effectively equal to the strongest baselines, which is odd if the point of SREV-Mol is to optimize for property satisfaction. This is important to clarify because the translation from query to SMARTS is fine-tuned in SREV-Mol and in no other baselines, so the gains of the method could be entirely due to this fine-tuning, which would render much of the components of the method redundant. The authors could run an ablation where the backbone model of SREV-Mol is fine-tuned without the additional modules to settle this question.

  • (minor) Some presentation-level fixes to improve the paper:

    • the fonts used in figures 1 and 2 should be increased to improve the reading experience: you can redesign the figures if needed to save space.
    • Section 3.3 introduces three acronyms in close succession (SER, AFT, AFE), which makes the paper dense to read, and the distinction between them is never stated plainly. I understood that SER is the model's output format, AFTs are the injected input vectors, and the AFE is the network that produces them. Could you remove some of them, or state the distinction explicitly? For instance I think naming the output format and the injected input vectors is unnecessary.
Questions for authors
  • Each swap's predicted property change (mean and standard deviation) comes from the matched pairs that support that swap. How many pairs does a swap need before its estimate is reliable, and how does performance change as you vary the support threshold (you set N_τ ≥ 10 in B.2) or the size of the pair database?

  • One swap is picked each time to maximize all properties at once, when in principle two combined edits together could form a better molecule overall, or we could pick a worse molecule temporarily to open up to better swaps later. Could you discuss how you can extend your module to account for these cases in the future, and how relevant they are in practice (i.e. how many of the failed runs needed combined edits)?

9 comments

We sincerely thank the reviewer for the constructive feedback and valuable suggestions for strengthening our experimental validation and presentation.

1. Learned Review vs. Simple Search Methods (W1)

The paper does not show that the learned Review step beats a simple search over the same candidate set.

→ To isolate the contribution of Review, we conducted an additional evaluation of Stage 2 on all 1,200 instances (900 ID and 300 Scaffold-OOD). Specifically, we provided identical seed SMILES and reference SMARTS to all methods and evaluated the optimization loop with a maximum of six edits (i.e., depth 6. Figure 3 (a) in the paper was measured at depth 5.). The results are shown in Table R1 below.

MethodSucc.↑(=Prop.↑)# Edits↓# Verify↓
Suggest Random57.17 (67.78 / 25.33)4.00 (3.62 / 5.14)5.00 (4.62 / 6.14)
Suggest Greedy61.92 (72.11 / 31.33)3.76 (3.38 / 4.91)4.76 (4.38 / 5.91)
Suggest Beam87.33 (95.00 / 64.33)2.31 (2.00 / 3.23)26.60 (21.1 / 43.0)
Suggest Exhaustive92.67 (97.11 / 79.33)2.36 (2.02 / 3.37)581.70 (258.0 / 1552.7)
Suggest Review (SREV-Mol)72.17 (82.78 / 40.33)3.41 (2.99 / 4.65)4.41 (3.99 / 5.65)

Table R1. Comparison with search-based methods under identical seeds. The values separated by "/" within the parentheses represent the ID and Scaffold-OOD scores, respectively.

Under identical seeds and a six-edit budget, SREV-Mol improves success rate(Succ.↑) over Suggest Greedy(Method for selecting the top candidate out of 4 candidates from MMPDB) by 10.25 percentage points (72.17% vs. 61.92%) while requiring slightly fewer verification calls (4.41 vs. 4.76). Beam and exhaustive search achieve higher success rates but require substantially more verification calls. Thus, learned Review provides a measurable advantage over a simple search baseline based on greedy edit selection at a comparable verification cost, although more extensive search strategies achieve higher success rates when additional computation is available.

We will include all of this content and analysis in the revised paper.

2. AFT Embeddings vs. Scalar AFE Scores (W2)

one ablation I suggest is to include the score output you get during pretraining (Eq. 11–13) in the prompt, instead of using the learned embedding vector.

→ As suggested by the reviewer, we constructed an additional variant of Review that replaces the AFT embeddings with scalar success scores predicted by the full pretrained AFE value-prediction network (Eqs. 11–13). These scores were inserted as text at the corresponding candidate positions in the Review input, and the model was fine-tuned using the same training data and settings.

Specifically, the Review input was modified as follows:

SREV-Mol (Original):

Edit 1: <|AFT1|>
Edit 2: <|AFT2|>
Edit 3: <|AFT3|>
Edit 4: <|AFT4|>
Out of range now: logD (above), ...

SREV-Mol (w/ AFE score text):

Edit 1: 0.62
Edit 2: 0.14
Edit 3: 0.09
Edit 4: 0.15
Out of range now: logD (above), ...

We first compared the score variant in the full end-to-end evaluation.

Succ.↑Struct.↑STRatio↑Prop.↑PTRatio↑Gain↑
IDSREV-Mol (w/ AFT score text)66.180.883.178.294.788.9
IDSREV-Mol67.882.885.280.696.690.5
OODSREV-Mol (w/ AFT score text)26.072.075.935.379.578.7
OODSREV-Mol29.071.376.238.081.982.8

Table R2. End-to-end comparison between AFT embeddings and textual AFE scores.

The results(Table R2) show comparable structural satisfaction between the two variants, while SREV-Mol achieves higher property satisfaction (Prop.↑ and PTRatio↑) and refinement gain on both datasets.

To further isolate the effect on property optimization, we conducted an additional controlled evaluation. Specifically, we fixed Stage 1, ground-truth SMARTS, and structurally valid seed SMILES, and evaluated only the Stage 2 optimization loop.

Succ.↑(=Prop.↑)# Edits↓# Verify↓
SREV-Mol (w/ AFT score text)70.92 (81.33 / 39.67)3.49 (3.07 / 4.73)4.49 (4.07 / 5.73)
SREV-Mol72.17 (82.78 / 40.33)3.41 (2.99 / 4.65)4.41 (3.99 / 5.65)

Table R3. Stage 2 optimization performance with AFT embeddings versus textual AFE scores. The values separated by "/" within the parentheses represent the ID and Scaffold-OOD scores, respectively.

The controlled Stage 2 evaluation shows a similar advantage for AFTs (72.17% vs. 70.92%) at comparable verification costs(Table R3). AFT-based methods appear to yield slightly better performance than simply providing a numerical score, likely because they offer a richer representation for each individual edit.

We will include these results in the revised paper.

3. Contribution of Review Beyond Structural Generation (W3)

Most of the empirical gains are on structural satisfaction when the method is supposed to be optimizing for properties. … The authors could run an ablation where the backbone model of SREV-Mol is fine-tuned without the additional modules to settle this question.

→ To examine whether Stage 2 contributes to property optimization beyond backbone fine-tuning, we trained an additional variant, SREV-Mol (w/o SER, w/o AFT), which directly selects an edit after Suggest without generating Review reasoning or using AFTs. This variant was fine-tuned using the same training data and settings.

As in the previous controlled evaluation, we fixed Stage 1 using ground-truth SMARTS, and structurally valid seed SMILES, and evaluated only Stage 2 with a maximum of six edits.

Succ.↑(=Prop.↑)# Edits↓# Verify↓
SREV-Mol (w/o SER, w/o AFT)69.753.514.51
SREV-Mol72.173.414.41

Table R4. Effect of the Review components on Stage 2 optimization.

Removing SER and AFTs reduces Stage 2 success from 72.17% to 69.75% under matched structural inputs(Table R4). Removing the Review (AFT+SER) component, a core element of Stage 2, results in a decrease of approximately 2.42 in Succ.↑(=Prop.↑), which demonstrates that performance gains do not stem solely from fine-tuning.

4. MMP Support Threshold Analysis

How many pairs does a swap need before its estimate is reliable, and how does performance change as you vary the support threshold ... or the size of the pair database?

→ The minimum support threshold Nmin⁡N_{\min} introduces a trade-off between the statistical support of property-change estimates and the coverage of available molecular transformations. Increasing Nmin⁡N_{\min} causes to have transformations supported by more matched pairs but reduces the number of available edits.

To examine this trade-off, we constructed MMP libraries with different support thresholds and evaluated their performance using search-based optimization methods.

Nmin⁡N_{\min}# Swaps in MMPDBSuggest Random (Succ.)Suggest Greedy (Succ.)Suggest Beam (Succ.)Suggest Exhaustive (Succ.)
6323,00767.772.694.997.6
10 (paper setting)139,83368.972.294.697.3
1485,48765.969.094.097.0
1860,51267.170.894.096.9
1006,65059.464.391.094.9
2002,58854.361.490.295.2
3001,42353.158.488.395.3

Table R5. Search success rates with different MMP support thresholds.

The results(Table R5) show that increasing Nmin⁡N_{\min} substantially reduces the library size(# Swaps in MMPDB**)**, generally leading to lower search success rates due to reduced edit coverage.

To further examine the effect of transformation support independently of library size, we sampled 1,423 swaps from each library and repeated the evaluation.

Nmin⁡N_{\min}# Swaps in MMPDBSuggest Random (Succ.)Suggest Greedy (Succ.)Suggest Beam (Succ.)Suggest Exhaustive (Succ.)
61,42320.628.650.060.6
101,42326.433.461.371.9
141,42332.138.872.981.6
181,42335.846.175.082.1
1001,42343.652.382.789.6
2001,42350.955.788.393.2
3001,42353.158.488.395.3

Table R6. Search success rates with a fixed number of swaps.

Together, these evaluations(Table R5, R6) reveal a trade-off between transformation support and edit-space coverage. A higher support count for a swap leads to more accurate edit recommendations from Suggest, thereby improving performance, but it also results in fewer usable swap rules, which lowers coverage.

Lowering Nmin⁡N_{\min} from 10 to 6 improves beam-search success by only 0.3 percentage points(Table R5) while doubling the average Suggest runtime (0.1 s to 0.2 s). We therefore retain Nmin⁡=10N_{\min}=10 as a practical balance between coverage, optimization performance, and computational efficiency.

5. Limitations in Handling Combined and Non-Monotonic Edits

Two combined edits together could form a better molecule overall, or we could pick a worse molecule temporarily to open up to better swaps later.

→ Applying multiple compatible swaps with low uncertainty could potentially improve refinement efficiency, although interactions between edits would need to be considered. Furthermore, an edit that temporarily reduces property satisfaction may create a molecular context that enables more favorable subsequent modifications.

One possible extension is to introduce reinforcement learning into Stage 2, allowing the editing policy to consider long-term property satisfaction rather than immediate improvements alone. We consider both combined-edit selection and long-term optimization promising directions for future work.

We will include these statements in the revised paper.

6. Figure Readability and SER/AFT/AFE Terminology

The fonts used in figures 1 and 2 should be increased" and "Section 3.3 introduces three acronyms in close succession (SER, AFT, AFE).

→ We will enlarge the fonts in Figures 1 and 2 and clarify the distinction among the three components in Section 3.3. Specifically, AFE extracts candidate representations, AFTs are the projected representations injected into the LLM, and SER is the structured reasoning format generated during Review. We will also include an annotated SREV-Mol trajectory in the appendix to illustrate how these components interact during inference.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers and Area Chair,

We sincerely thank you for the time and effort devoted to reviewing our manuscript.

As highlighted by the reviewers, our work addresses a practically relevant molecular constraint-satisfaction problem (Reviewers ANzC, jnxf), introduces a principled and carefully designed framework with structure-preserving edit-space reduction and iterative refinement (All Reviewers), and demonstrates strong empirical performance on both ID and Scaffold-OOD benchmarks (Reviewers azVz, jnxf), even with a relatively small 8B backbone (Reviewer azVz).

We greatly appreciate the constructive feedback and suggestions. In response, we have carefully revised the manuscript and conducted additional experiments and analyses to address the reviewers' concerns, as summarized below:

  • Reviewer ANzC
    • Learned Review vs. Simple Search Methods (W1) [Section 4.3.2, Table 5]
    • AFT Embeddings vs. Scalar AFE Scores (W2) [Section F.3, Table 14]
    • Contribution of Review Beyond Structural Generation (W3) [Section 4.3.2, Table 5]
    • MMP Support Threshold Analysis [Section F, Table 11]
    • Limitations in Handling Combined and Non-Monotonic Edits [Section 5]
    • Figure Readability and SER/AFT/AFE Terminology [Figure 1, Section 3.3, Figure 11]
  • Reviewer azVz
    • Synthetic Description Quality and Human-Written Generalization [Section F.5, Table 15, 16]
    • Scaffold Similarity and OOD Generalization [Section F.4, Figure 14]
    • Inference Cost and Reasoning-Budget Sensitivity [Section F.6, Table 18, Figure 15]
    • Chemical Correctness of Review Reasoning [Section C.3]
  • Reviewer jnxf
    • Independent Contributions of SER and AFTs Beyond Stage 1 (W1, Q1) [Section 4.3.2, Table 5]
    • Task-Aligned Optimization Baselines (W2, Q3) [Section F.7, Table 19]
    • Clarification of Excluded Chemistry Agent Tools [Section E.1]
    • Train-Test Independence in Edit-Selection Evaluation (W2, Q2) [Figure 3(b), Section E.4]
    • Effect of Round-Trip Verification Filtering (W2) [Section F.5, Table 17]
    • Performance by Number of Property Constraints (W2, Q3) [Section 4.3.2, Table 6]
    • Statistical Significance of Scaffold-OOD Differences (W2, Q3) [Section E.2, Figure 12]

The newly added analyses and major revisions are highlighted in blue\textcolor{blue}{\text{blue}} in the revised manuscript for ease of review.

We hope that our detailed responses and revisions adequately address the reviewers' concerns and clarify the contributions and limitations of our work.

Thank you again for your valuable feedback and consideration.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.