acceptodds
Under review as a conference paper at ICLR 2027

Solvent: Long-Horizon Reasoning for Multi-Step Retrosynthesis

27 views

Abstract

Multi-step retrosynthesis requires recursively transforming a target molecule through successive reaction steps into purchasable building blocks, and recent work has begun to draw on large language model (LLM) reasoning to tackle the task. In its latest evolution, LLM reasoning has entered the long-horizon regime: models pursue objectives across interdependent steps, interact with environments, and adapt to feedback. This shift has markedly advanced LLM capabilities, as illustrated by breakthroughs in AI-driven mathematical research. At this frontier, we introduce Solvent, an 8B LLM that successfully extends long-horizon reasoning to multi-step retrosynthesis. Through a two-round post-training curriculum, Solvent learns to interleave iterative synthesis planning, external availability verification, and adaptive route revision. Experiments show that, following a performance boost from RL, Solvent scores 61.86% across four benchmarks, surpassing mid-scale language models and prior search-based retrosynthesis systems, and rivaling frontier trillion-parameter LLMs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multi-step retrosynthesis, Long-horizon reasoning, LLM agents, Reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nQj2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SOLVENT, a 8B LLM trained on top of R^3-pretrained 8B model with 5 training phases for direct multistep retrosynthesis route generation without the need of a single-step model. SOLVENT outperforms all LLM based methods with 8B parameters. SOLVENT also outperforms conventional search-based methods when comparing LLM call budgets against single-step model call budgets. Additionally, SOLVENT is shown to perform long-horizon reasoning since the performance increases as the feedback turn proceeds.

Critical strengths
  1. Competitive Performance SOLVENT achieve better performance compared to other LLM baselines on multiple benchmarks with 8B of parameters. When constraining the number of single-step model calls in search-based baselines the same as the number of LLM calls in SOLVENT, SOLVENT also outperforms conventional search-based algorithms.

  2. Comprehensive Experiments SOLVENT conducts detailed ablation experiments with detailed description about the experiment setup in the appendix.

Critical weaknesses
  1. Limited Chemical Viability The primary success criterion provides limited evidence of chemical feasibility, as template substructure support and atom-source accounting do not establish that the proposed transformations produce the intended products under compatible conditions. Though additional analysis on route quality judged by ChemSensor, SAScoure and round trip validity is provided in Table 21, it's limited to 400 shared targets and SOLVENT failed to outperform KimiK3 and other proprietary LLMs in terms of round trip validity.

  2. Over-claimed Superiority The search-budget comparison uses different accounting units for Solvent and conventional planners (number of LLM calls for SOLVENT or other LLM based methods and number of single-step model calls for conventional planners). Consequently, the reported advantage at equal numerical budgets does not establish superior computational or search efficiency. Moreover, route quality metrics in Table 21 show that SOLVENT provide worse routes compared to other LLMs. While the training stage ablation is provided, the 5-stage training pipeline is still a sign of over-engineering compared to conventional planners with no clear advantage. The "long-horizon reasoning" capability is shown via the increasing performance as the turn goes up, but it's still unclear whether the historical interaction played important role in its reasoning or just added noise to the current iteration.

Questions for authors
  1. Is SOLVENT really history-dependent, or can SOLVENT really do long-horizon reasoning? Currently it's shown that the performance (in terms of success rate) increases as the turns goes up while maintaining all interactions in the context, what if only the latest turn is kept in the context? How many of prior history are kept due to the context window limitation?
  2. Could you provide extended ChemSensor, SAScore, Round-trip evaluation to other experiments?
  3. Can SOLVENT still outperform conventional baselines or other LLM-baselines with single-step models if the compute is the same?
  4. The design of the current 5-stage training pipeline is not intuitive, could you share more about the design choices and why the starting point (R3R^3 pretrained Qwen3-8B) performs differently with R3R^3 (0.18% of Initial Backbone in table 3 v.s. 19.03% of R3 in table 1)?
8 comments
AuthorsOpenReview ↗

Response to Reviewer nQj2

We thank the reviewer for the detailed feedback and suggestions for additional evaluation.

Weakness 1 & Weakness 2

Template-substructure support and atom-source accounting do not establish that the proposed transformations produce the intended products under compatible conditions. The route-quality analysis is limited to 400 shared targets, and Solvent does not outperform Kimi K3 and other proprietary LLMs in round-trip validity. Does comparing LLM calls with single-step model calls in Table 2 establish superior search or computational efficiency?

Solvent's template-free generation allows joint planning of complete routes and exploration of new substrate combinations, supported by additional chemical checks. Template-based planners constrain reaction proposals through template application. These checks assess chemical plausibility, with compatible reaction conditions requiring further expert and experimental validation. Forward round-trip evaluation also depends on a learned predictor and can produce both false positives and false negatives [1]. We therefore use complementary diagnostics to assess different aspects of route quality.

The route-quality results show a mixed profile, with competitive chemical-support metrics and room to improve forward-model consistency. Table 23 uses the same fixed 400 targets for all models. Solvent has lower round-trip recovery than Kimi K3 (9.02% vs. 13.44%), alongside stronger synthetic-accessibility improvement and ChemCensor reaction-center and functional-group support.

Our budget alignment is designed to compare search efficiency across different planning methods. For Table 2, Solvent's search budget N counts the cumulative number of reaction steps in submitted routes, accounting for the reaction-level work used to find a complete route. We have clarified this definition and comparison scope in Appendix C.2.

Our contribution is a practical, reproducible post-training recipe that develops long-horizon retrosynthetic reasoning in an 8B model to rival frontier LLMs on this task. Kimi K3 is a strong reference model whose post-training task synthesis includes chemistry [2]. The experiments explain how behavior stabilization and controlled bootstrapping depth prepare the model for effective RL. We will release the stage-wise checkpoints and complete training pipeline to help the community reproduce and build on these findings. We have clarified the training rationale in Section 4. Appendix F.2.4 presents the complementary route-quality comparisons.

Question 1

Does Solvent use interaction history for long-horizon reasoning? What happens if only the latest turn is kept in context, and how much prior history is retained under the context-window limit?

Solvent retains the complete interaction history within its 32k-token context window. We tested the suggested latest-turn-only setting, retaining the fixed prompt and the most recent model response with its corresponding tool feedback. The comparison covers all 1,440 targets with up to five interaction turns per trajectory:

BenchmarkTargetsFull-history success (%)Latest-turn-only success (%)
Pistachio Hard10062.0047.00
Pistachio Reachable15078.0070.00
USPTO-19019048.9548.95
ChEMBL-10001,00058.5057.00
Macro average1,44061.8655.74

The 6.12 percentage-point drop in macro-average success supports the value of earlier interactions for route revision. We have added this ablation in Appendix F.1.6 (Table 18) and two case studies in Appendix G.4 showing how Solvent uses earlier interaction history to guide route revision.

Question 2

Could you provide extended ChemSensor, SAScore, Round-trip evaluation to other experiments?

We extended the ChemCensor, SA score, and forward round-trip evaluations to RETRO-R1 and EG-MCTS from Table 2, using the same shared 400-target subset and evaluation configurations as Table 23. The results show complementary strengths across methods.

MethodTarget–leaf ΔSASA-reducing steps (%)RC routes/100 stepsFG routes/100 stepsStock-complete & route RT@20 (%)
RETRO-R10.36267.1714.776.7510.63
EG-MCTS0.56071.8018.128.0816.01
Solvent0.82784.698.624.119.02

Solvent achieves the largest target–leaf ΔSA (0.827) and the highest proportion of SA-reducing steps (84.69%). Solvent's template-free generation allows flexible reaction proposals, while verifying chemical feasibility remains a challenge for template-free methods. In this comparison, RETRO-R1 and EG-MCTS obtain higher ChemCensor RC/FG and round-trip scores. We have added the expanded evaluation and discussion in Appendix F.2.4 (Table 25).

Question 3

Can SOLVENT still outperform conventional baselines or other LLM-baselines with single-step models if the compute is the same?

To address the question about computational efficiency, we complemented the search-budget results in Table 2 with runtime measurements of Solvent and RETRO-R1 on all 1,440 targets under matched hardware and inference configurations. As discussed in Appendix C.2, computational effort is difficult to align across LLM generation, single-step model prediction, and tree search.

Solvent has lower mean per-target termination times than RETRO-R1 across all four benchmarks. The table reports mean termination time and mean per-target effective output rate, including tool calls.

BenchmarkSolvent time (s)RETRO-R1 time (s)Solvent effective tokens/sRETRO-R1 effective tokens/s
USPTO-190741.891,845.8211.450.77
ChEMBL-1000551.801,369.5914.240.77
Pistachio Reachable462.76763.4011.440.78
Pistachio Hard718.181,532.1511.300.78

Solvent's higher token usage is compatible with lower overall planning time. For an LLM planner coupled with a single-step model, the total cost includes LLM generation, single-step prediction, and search. Solvent directly generates and revises complete routes.

Solvent uses more extensive reasoning to improve route planning, naturally requiring additional tokens. It allocates computation to global synthesis strategy, precursor availability, and feedback-guided route revision rather than explicit search over single-step disconnections. This route-level reasoning supports the higher success reported in Table 2 under the stated search budgets. We have added the runtime measurements in Appendix F.2.3 (Table 22) and clarified the scope of budget alignment in Appendix C.2.

Question 4

What motivates the five-stage training pipeline? Why does the Initial Backbone (pretrained Qwen3-8B) achieve 0.18% in Table 3, compared with 19.03% for R3 in Table 1?

We designed our two-round curriculum based on the insight that stable behavior and sufficient exploration capacity are needed for effective RL. Round 1 establishes the foundation for subsequent capability improvement. Warm-up builds retrosynthetic knowledge, teacher SFT introduces multi-turn planning, and RL consolidates the model's own successful behaviors into a stable seed policy, addressing malformed queries and truncated responses. Round 2 builds on this foundation to prepare a stronger starting policy for final-stage RL. We bootstrap the warmed-up policy using successful trajectories from the seed policy and control bootstrapping depth to retain sufficient exploration capacity (Appendix F.1.2, Figure 6). Applying the same final-stage RL directly to the teacher-SFT policy yields substantially lower validation success, supporting the importance of this preparation (Appendix F.1.3, Figure 7). The five stages thus progress from behavioral stabilization to task-performance improvement. We have clarified this rationale in Section 4.

The 0.18% and 19.03% results correspond to different checkpoints from the R3 work [3]. The Initial Backbone in Table 3 (0.18%) is the Qwen3-8B-Base checkpoint after R3's continued pretraining (CPT), which we use to initialize Solvent. The R3 baseline in Table 1 (19.03%) is their final model after the complete post-training procedure. We have clarified these checkpoint identities in Section 5.1 and the Table 3 caption.

[1] Choe, J., Chen, S., and Jung, Y. Assessing the extrapolation capability of template-free retrosynthesis models. Digital Discovery, 2026. https://doi.org/10.1039/D6DD00072J

[2] Kimi Team. Kimi K3: Open Frontier Intelligence. 2026. Section 4.2.2, Figure 9. https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf

[3] Wang et al. R3: End-to-End Reasoning-based Planning for Multi-step Retrosynthesis via Reinforcement Learning. ACL 2026. https://aclanthology.org/2026.acl-long.1745/

Reviewer_xj2MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Solvent is an 8B LLM for multi-step retrosynthesis that generates complete routes and revises them using building-block availability feedback. Starting from a retrosynthesis-pretrained checkpoint, it uses a two-round pipeline combining supervised fine-tuning, trajectory bootstrapping, and reinforcement learning.

Critical strengths

The paper provides evidence that an 8B model can benefit from inventory-guided route revision. On the shared 400-target subset, Solvent achieves a 63.26% macro-average success rate, close to the strongest evaluated models. The paired feedback intervention also shows gains from true versus shuffled or masked inventory feedback.

Critical weaknesses
  1. Limited incremental contribution. Complete-route generation, RL-trained retrosynthesis agents, teacher distillation, and trajectory bootstrapping are established ideas. Solvent's main distinction is to combine direct reaction generation with multi-turn inventory-guided revision, rather than relying on an external single-step model. The reported improvement over R3 does not clearly separate the contribution of the proposed design from the effects of additional supervision, training effort, and adaptation to inventory interaction.

  2. Checkpoint provenance and the relationship to the R3 baseline need clarification. Appendix E states that Solvent starts from the retrosynthesis-pretrained Qwen3-8B checkpoint developed for R3. The relationship between this checkpoint, the R3 baseline in Table 1, and the Initial Backbone in Table 3 could be clarified. For example, R3 achieves 19.00% on Pistachio Hard, whereas the Initial Backbone achieves 0.00%. I am also unable to locate a public release of the relevant R3 weights and training artifacts. The supplementary code does not specify how to access the checkpoint, which makes it difficult to assess the baseline comparison and independently reproduce the results. The planned release of Solvent upon acceptance also leaves access to these underlying checkpoints and data unclear.

Questions for authors
  1. Could the authors clarify the specific advance beyond combining complete-route generation with inventory feedback and established post-training techniques? Which controlled comparisons distinguish this contribution from the effects of additional training data, training effort, and tool adaptation?

  2. Please identify the exact checkpoints used for the R3 baseline, the Initial Backbone, and Solvent's initialization, including their completed training stages. What accounts for the performance difference between Tables 1 and 3?

  3. How are the R3-related checkpoints and route corpus obtained? Can the authors provide an anonymous access point or a reproducible acquisition procedure? Does the planned release cover the initialization checkpoint and necessary training artifacts, or only the final Solvent weights and code?

4 comments
AuthorsOpenReview ↗

Response to Reviewer xj2M

We appreciate the reviewer's helpful comments on the contribution and reproducibility of our work.

Weakness 1 & Question 1

Could the authors clarify the specific advance beyond combining complete-route generation with inventory feedback and established post-training techniques? Which controlled comparisons distinguish this contribution from the effects of additional training data, training effort, and tool adaptation?

To assess the training design and feedback-guided revision, we present curriculum and feedback ablations and add a comparison with R3 Pass@5 under a shared limit of five route-generation attempts per target.

Our contribution is a post-training recipe that develops long-horizon retrosynthetic reasoning in an 8B model to rival frontier LLMs on this task. Our curriculum develops task capability that is largely absent from the initial backbone. Kimi K3, a frontier reference model, includes chemistry in its post-training task synthesis [1]. Our initial backbone achieves 0.18% average success, increasing to 61.86% after our curriculum (Table 3). The shared 400-target comparison shows competitive performance against frontier LLMs (Appendix C.2, Table 6).

Stable behavior and sufficient exploration capacity prepare the model for effective RL. We designed our two-round curriculum based on this insight, first stabilizing interaction behavior and then improving task performance through controlled bootstrapping and RL. Our ablations support this design and the role of informative feedback in route revision (Appendices F.1.2–F.1.3, Figures 6–7, and Appendix F.1.5, Table 17). We have further clarified the training rationale in Section 4.

Solvent outperforms R3 Pass@5 on all four benchmarks, with a 19.43 percentage-point gain in macro-average success. R3 [2] generates five independent routes using its official inference procedure, while Solvent generates an initial route followed by up to four revisions. The comparison covers all 1,440 targets under the same inventory and chemical checks (Appendix F.2.7, Table 27).

BenchmarkTargetsR3 Pass@5 (%)Solvent final-route success (%)
Pistachio Hard10038.0062.00
Pistachio Reachable15058.6778.00
USPTO-19019036.8448.95
ChEMBL-10001,00036.2058.50
Macro average1,44042.4361.86

[1] Kimi Team. Kimi K3: Open Frontier Intelligence. 2026. Section 4.2.2, Figure 9. https://arxiv.org/html/2607.24653v2#S4.SS2.SSS2

[2] Wang et al. R3: End-to-End Reasoning-based Planning for Multi-step Retrosynthesis via Reinforcement Learning. ACL 2026. https://aclanthology.org/2026.acl-long.1745/

Weakness 2 & Question 2

Please identify the exact checkpoints used for the R3 baseline, the Initial Backbone, and Solvent's initialization, including their completed training stages. What accounts for the performance difference between Tables 1 and 3?

The R3 baseline in Table 1 uses the final R3 checkpoint after its complete post-training procedure. The Initial Backbone in Table 3 is the intermediate Qwen3-8B-Base checkpoint after R3's continued pretraining (CPT). Solvent is initialized from this same CPT checkpoint, then follows our five-stage post-training curriculum (Appendix E).

The 19.00% versus 0.00% success rates on Pistachio Hard therefore correspond to the fully trained R3 model and its CPT intermediate checkpoint, respectively. We have clarified the baseline checkpoint in Section 5.1 and the initialization checkpoint in the Table 3 caption.

Question 3

How are the R3-related checkpoints and route corpus obtained? Can the authors provide an anonymous access point or a reproducible acquisition procedure? Does the planned release cover the initialization checkpoint and necessary training artifacts, or only the final Solvent weights and code?

R3's CPT checkpoint, final model, and SFT route corpus were publicly released by the R3 authors before our submission:

R3 checkpoint after continued pretraining, used to initialize Solvent: https://modelscope.cn/models/wyifei/Reinforced-Reasoning-Retrosynthesis-CPT

Fully trained R3 checkpoint, used as the baseline in Table 1: https://modelscope.cn/models/wyifei/Reinforced-Reasoning-Retrosynthesis-RL

R3 route corpus for supervised fine-tuning (SFT) and associated training code: https://gitee.com/wyifei26/r3-llamafactory

Solvent's warm-up data are derived from this route corpus through deduplication, benchmark decontamination, and Intern-S1 re-annotation with quality filtering, as detailed in Appendix D.1. The supplementary material specifies the source files, resource versions, and preprocessing steps needed to reproduce data acquisition and preparation.

Our release will cover the initialization checkpoint, checkpoints from all post-training stages, and the complete training code, configurations, and data-processing scripts, enabling reproduction of the full training pipeline and intermediate results.

AuthorsOpenReview ↗

Follow-up on our response

We wanted to briefly follow up on our response and ask whether any questions remain. We would be happy to provide further clarification. Thank you for your consideration.

Thank you for the response. After reviewing the discussion and other reviews, I am inclined to maintain my rating.

Reviewer_pDnaOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Solvent, an 8B LLM for multi-step retrosynthesis through iterative planning, inventory checking and route revision. A two-round curriculum combines supervised fine-tuning, reasoning bootstrapping and reinforcement learning. Experiments report competitive performance against frontier LLMs and improvements over previous search-based planners.

Critical strengths
  1. The results are promising for an 8B model. The comparison on the same 400 targets supports its competitive performance against frontier LLMs and substantial improvement over R3.

  2. The analysis goes beyond reporting final performance. Training-stage comparisons and the paired feedback intervention provide useful evidence for the curriculum and adaptive revision, especially the clear gains from true inventory feedback.

Critical weaknesses
  1. The primary chemical-validity criterion needs clearer justification. Template-substructure matching and element counts do not by themselves establish the exact reaction transformation. Since this verifier also supplies RL rewards, its false-acceptance behavior matters.

  2. The search comparison is not sufficiently controlled (Table 2). The budget counts single-step model calls, route-level operations and Solvent’s submitted reaction steps for different methods. These counts do not establish equivalent search effort. It is also unclear whether baseline routes use the same inventory release and chemical checks. The superiority claim should be qualified unless these conditions are aligned.

Questions for authors
  1. Does the ASKCOS template check require exact product reconstruction from the proposed precursors, or only matching substructures? Please clarify connectivity and stereochemistry checks, and any existing audit of false acceptances.

  2. See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer pDna

We thank the reviewer for the careful review and constructive feedback.

Weakness 1 & Question 1

Does the ASKCOS template check require exact product reconstruction from the proposed precursors, or only matching substructures? Please clarify connectivity and stereochemistry checks, and any existing audit of false acceptances. How is this chemical-validity criterion justified, given that it also supplies RL rewards?

Our ASKCOS check uses substructure matching on both sides of the same reaction template, together with atom-source accounting (Appendix C.1). It checks connectivity within matched substructures and requires a connected, acyclic route, without exact product reconstruction. Inventory lookup preserves stereochemistry through canonical isomeric SMILES.

Our design aims to preserve exploration of new substrate combinations and complete routes beyond an external single-step model's candidates. Reliable verification remains a shared challenge in template-free retrosynthesis, with false positives and false negatives also documented for forward round-trip evaluation [1].

Independent diagnostics provide empirical support for our verification criterion. Appendix F.2.4 (Table 24) shows stronger ChemCensor support and forward-model recovery for accepted routes within the same stock-complete set. We have clarified the verification scope and rationale in Appendix C.1.

Weakness 2 & Question 2

Do the different budget units in Table 2 establish comparable search effort? Do baseline routes use the same inventory release and chemical checks? The superiority claim should be qualified unless these conditions are aligned.

To address the concerns about inventory consistency and search effort, we evaluated Solvent with eMolecules (2019) and measured runtime against RETRO-R1 under matched hardware and inference configurations. We also clarified the chemical checks and search-budget definitions used for each method.

Solvent maintains high success rates with eMolecules (2019) across all four benchmarks. We evaluated all 1,440 targets at N=100 using the same eMolecules (2019) inventory as Retro* and RETRO-R1. The results remain comparable to those obtained with eMolecules (2026), supporting robustness to the inventory release (Appendix F.2.6, Table 26). The table reports success rates (%).

BenchmarkTargetseMolecules (2026)eMolecules (2019)
USPTO-19019087.3690.00
ChEMBL-10001,00093.2092.70
Pistachio Reachable15099.3399.33
Pistachio Hard10093.0093.00

Solvent's template-free generation motivates additional chemical screening during training and evaluation. Template-based planners constrain reaction proposals through template application and retain their evaluation protocols. Solvent's outputs additionally undergo ASKCOS template-support and atom-source checks. We have clarified these protocols in Appendix C.2.

Our budget alignment is designed to compare search efficiency across planning methods. For Solvent, we count the cumulative reaction steps in submitted routes, reflecting the reaction-level work involved in constructing candidate routes. We have clarified this definition and comparison scope in Appendix C.2.

Solvent has lower mean per-target termination times than RETRO-R1 on all four benchmarks under matched hardware and inference configurations. The table reports mean termination time and mean per-target effective output rate over all 1,440 targets, with elapsed time including tool calls. Each method retains its own stopping criterion (Appendix F.2.3, Table 22).

BenchmarkSolvent time (s)RETRO-R1 time (s)Solvent effective tokens/sRETRO-R1 effective tokens/s
USPTO-190741.891,845.8211.450.77
ChEMBL-1000551.801,369.5914.240.77
Pistachio Reachable462.76763.4011.440.78
Pistachio Hard718.181,532.1511.300.78

[1] Choe, J., Chen, S., and Jung, Y. Assessing the extrapolation capability of template-free retrosynthesis models. Digital Discovery, 2026. https://doi.org/10.1039/D6DD00072J

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your time and feedback. We would appreciate hearing whether our response has addressed your concerns. Please let us know if anything remains unclear.

Comments on the paper

5 comments
AuthorsOpenReview ↗

General response

We thank all reviewers for their careful reading and constructive feedback. We have revised the manuscript, with changes highlighted in dark blue.

To examine the role of interaction history, we added a latest-turn-only ablation and two case studies showing how earlier observations inform later revisions (Appendices F.1.6 and G.4). The decrease in success when earlier history is removed supports its value for route revision. We also compared Solvent with R3 Pass@5 under a shared limit of five route-generation attempts per target (Appendix F.2.7). Solvent achieves higher success on all four benchmarks.

To strengthen the evaluation, we compared results using eMolecules (2019) and eMolecules (2026) and measured runtime against RETRO-R1 under matched hardware and inference configurations, examining inventory sensitivity and practical computational cost (Appendices F.2.6 and F.2.3). We further added route-quality assessments against RETRO-R1 and EG-MCTS to compare complementary aspects of chemical quality (Appendix F.2.4).

We have clarified the chemical-verification scope and search-budget definition (Appendices C.1–C.2), the checkpoint identities (Section 5.1 and Table 3), and the rationale for our two-round curriculum: stable interaction behavior and sufficient exploration capacity support effective downstream RL (Section 4). Detailed responses to each reviewer accompany these revisions.

The paper provides target-level decontamination statistics for its post-training data. However, Solvent is initialized from the retrosynthesis-pretrained R3 checkpoint, whose continued pretraining includes USPTO-Full reaction data. Since USPTO-190 is also derived from USPTO-Full, it remains unclear whether benchmark reactions were already encountered during initialization. The authors should clarify whether the inherited checkpoint was decontaminated against all four evaluation benchmarks and provide reaction-level overlap statistics, including overlaps involving intermediate molecules in reference routes. Without this analysis, the reported zero target overlap does not fully establish training–evaluation separation.

AuthorsOpenReview ↗

Response to the public comment on training-data separation

Thank you for raising this question.

We applied strict deduplication and decontamination throughout Solvent's post-training and verified substantial structural separation from the evaluation targets. Beyond zero target overlap with all 1,440 evaluation targets, Table 5 shows that 81.49% of scaffold-bearing test targets have Murcko scaffolds absent from the post-training data, and only 5.28% of test targets have a training neighbor with Morgan similarity ≥0.8.

R3's continued pretraining corpus has not been released, and we do not have access to it. We therefore cannot verify whether it was decontaminated against commonly used benchmarks, including our four test sets, or provide reaction- and intermediate-level overlap statistics.

We have also examined the inherited CPT checkpoint on all four multi-step retrosynthesis benchmarks and found negligible success. Table 3 reports:

CheckpointPistachio Hard (%)Pistachio Reachable (%)USPTO-190 (%)ChEMBL-1000 (%)Macro average (%)
R3 checkpoint after continued pretraining0.000.000.530.200.18

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.