acceptodds
Desk rejected at ICLR 2027

RetroRoute: Route-Level Agentic Decision Making for Multi-Step Retrosynthesis

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

18 views

Abstract

Development of robust and effective strategies for retrosynthetic planning requires effective decision making throughout of the synthesis process. A critical step in achieving this goal is enabling an agent to evaluate intermediate choices according to their long-term contribution to successful route completion. Current machine learning-based methods often focus on single-step prediction or heuristic search, optimizing locally plausible reactions without explicitly learning how sequential decisions affect the final synthetic route, resulting in suboptimal planning for long and complex syntheses. Therefore, we introduce RetroRoute, an advanced agentic framework for multi-step retrosynthesis that optimizes decisions over the entire synthesis route. The proposed LLM-based REVIEWER and state-aware ANALYZER form the RetroRoute agent for candidate generation and route-level decision making. First, the planning-conditioned REVIEWER generates and reviews candidate reactions based on the target, current intermediate, and search budget, with route-aware utility refinement further identifying chemically feasible and route-compatible alternatives. Furthermore, the state-aware ANALYZER represents each search state with search history and learns to rank candidate reactions toward successful route completion. As a result, RetroRoute optimizes sequential decisions for long-term route utility, enabling agentic global planning toward synthesis completion. In experiments on the USPTO pathways dataset RetroBench, RetroRoute outperforms state-of-the-art methods, achieving up to a 3% improvement in Top-1 test accuracy, particularly for long synthetic routes. These results demonstrate the superiority of RetroRoute in using route-level information to guide agentic decisions over the complete retrosynthetic planning process. They also demonstrate its potential for advancing autonomous retrosynthetic planning and facilitating reliable synthesis route design. Code is available at https://github.com/Su-yq/RetroRoute.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Retrosynthesis Planning, Large Language Models, Global Planning, Decision Making, Self-Enhancement

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_PmE8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RetroRoute, a multistep retrosynthesis framework combining a planning-conditioned MolT5 proposal model, utility-filtered positive refinement, and a ChemDFM-based state-aware listwise reranker. The method achieves improved Top-1 reference-route recovery on RetroBench, with particularly strong gains over proposal-only variants under greedy search.

Critical strengths
  1. The framework addresses a relevant limitation of single-step retrosynthesis: a locally plausible reaction may be unhelpful for completing a synthesis route.

  2. The controlled experiments show consistent gains from the planning scaffold, URPO, and the trained Analyzer. The greedy-search experiments are particularly useful in demonstrating improved first-choice branch ordering.

Critical weaknesses
  1. The current evidence does not convincingly establish that the improvement comes from route-level or state-aware decision making. The Analyzer receives not only route-level information such as the frontier and history, but also strong local candidate features including proposal rank, structural validity, and purchasable-reactant count. The comparison against an unadapted ChemDFM therefore does not isolate the contribution of the evolving search state. A trained local reranker with the same candidate-level features but without route history/frontier would be a much more informative control.

  2. The evaluation is strongly tied to reference-route recovery, making the claimed improvement in general route quality unclear. The main “route exact match” metric is based on matching the terminal starting-material set to a recorded reference, while the training procedure also uses route annotations and a reference-alignment utility term. It is therefore difficult to determine whether RetroRoute learns general long-term synthesis utility or primarily becomes better at recovering benchmark-compatible reference routes.

Questions for authors
  1. Can the authors provide an ablation that removes route-level state information (e.g., frontier, history, and remaining budget) while retaining the same trained reranker and local candidate features?
  2. How exactly are the positive actions PtP_t used to train the Analyzer constructed? In particular, are they defined by reference-route membership, URPO utility, downstream search success, or some combination of these signals?
  3. Can the authors report reference-independent measures of route quality or chemical validity?
  4. What is the contribution of the individual URPO utility terms, especially urouteu_{route} and urefu_{ref}?
  5. Can a smaller non-LLM reranker trained with the same listwise supervision achieve comparable performance?
5 comments
AuthorsOpenReview ↗

Thanks for the review - Weakness1

Dear Reviewer PmE8, We sincerely appreciate the time and effort you have taken to provide valuable feedback on our work. Your insightful comments have been immensely helpful in improving the clarity and quality of our manuscript. We have carefully considered each of your suggestions and have made the appropriate revisions to address your concerns. Below, we present detailed, point-by-point responses to your observations.

Response to Weakness1:

We agree that the original comparison with an unadapted ChemDFM does not fully isolate the contribution of route-state information. We therefore provide an additional controlled ablation and further clarify the motivation behind our framework.

First, the central motivation of RetroRoute is to bridge the gap between local reaction plausibility and sequential route-level decision making. Local descriptors, such as proposal rank, structural validity, and purchasable-reactant count, provide valuable evidence about individual reactions. However, they do not fully characterize the evolving synthesis problem, where the suitability of a reaction also depends on unresolved intermediates, previous decisions, and the remaining search budget. Accordingly, our contribution is not simply to introduce additional features into a reranker, but to establish a unified planning framework that separates candidate generation from state-conditioned branch evaluation. The Reviewer expands the candidate space through planning-conditioned generation and URPO refinement, while the Analyzer integrates local reaction evidence with the evolving search context to guide branch selection.

Second, we conduct a controlled ablation to explicitly assess the contribution of route-state information. As detailed in our response to Question 1, the Local-only ChemDFM retains the same pretrained initialization, LoRA adaptation, training examples, listwise supervision, and candidate-level descriptors as the full Analyzer, but excludes the unresolved frontier, route history, and remaining search budget.

Table R3-1: Contribution of route-state information in the Analyzer.

AnalyzerLocal featuresRoute-state featuresBeam Top-1Beam Top-3Beam Top-5Greedy DFS Top-1
Local-only ChemDFMYesNo36.6%46.3%49.8%35.52%
Full-state ChemDFMYesYes37.6%46.7%50.1%36.67%

Table R3-1 indicates that local descriptors already provide a strong foundation for candidate evaluation, while route-state information contributes additional, complementary planning context. We therefore emphasize that the advantage of RetroRoute arises from integrating complementary reaction-level and route-level information within a sequential decision-making framework, rather than attributing the overall improvement to any single input component. This distinction is central to our methodological contribution and is further supported by the controlled ablation.

AuthorsOpenReview ↗

Thanks for the review - Weakness2

Response to Weakness2:

We sincerely thank the reviewer for raising this important concern. We agree that reference-route exact match alone cannot establish general synthesis quality, particularly when route annotations are used during training. We would therefore like to clarify the distinction between reference-derived training supervision and reference-independent route evaluation, and provide additional evidence beyond reference-route recovery.

1. The role of reference information in training and inference. We acknowledge that the Analyzer uses reference-reaction supervision at teacher-forced training states, and that URPO incorporates a reference-alignment utility signal. These annotations provide supervision for learning candidate preferences, while inference proceeds using model-generated candidates and the dynamically evolving search state.

2. RetroRoute explicitly connects sequential reaction decisions through iterative route-context updates. At each planning step, the search state incorporates the unresolved frontier, accumulated route history, and remaining search budget. After a reaction is selected, the controller updates the frontier and records the decision in the route history. The Analyzer then evaluates subsequent candidates using this updated context. Consequently, later decisions are conditioned on the partial synthesis route constructed by preceding model decisions, rather than being independently matched against a fixed reference trajectory. This iterative mechanism is central to our route-level planning design.

3. Additional reference-independent evaluation metrics. We evaluate the quality of routes generated by RetroRoute on 5,838 test targets from multiple reference-independent perspectives, including route diversity, average number of reaction steps, and route complexity, as summarized in Table R3-3. Overall, these metrics indicate that RetroRoute produces diverse, structurally consistent, and moderately complex routes, providing complementary evidence of route generation quality beyond reference-route matching.

Table R3-3. Reference-independent evaluation of generated routes under beam search.

MetricResult
Completed targets5,826 / 5,838
Beam search completion rate99.79%
Top-5 reaction diversity0.791
Top-5 starting-material diversity0.581
Distinct reaction routes (Top-5)4.91
Average reaction steps (Top-1)3.13
Average maximum route depth (Top-1)3.02
Average distinct starting materials (Top-1)3.46
Valid SMILES rate (Top-1)100%
Acyclic decomposition graph rate (Top-1)99.47%

Overall, we position RetroRoute as a framework for state-aware sequential retrosynthetic planning. Importantly, RetroRoute is not directly trained using rollout-derived route-completion rewards. Instead, URPO uses predefined chemical and route-related utility signals as offline reward surrogates, while the Analyzer learns state-conditioned candidate preferences through teacher-forced reference-action supervision. Our contribution therefore lies in integrating these complementary supervision signals with dynamically updated planning states, rather than explicitly learning a calibrated long-horizon return from observed synthesis outcomes. We hope this clarification addresses the reviewer's concern about the distinction between reference-derived training supervision and reference-independent route evaluation, and shows that RetroRoute is not merely recovering reference routes but performing state-aware sequential planning.

AuthorsOpenReview ↗

Thanks for the review - Question1 & 2

Response to Question1:

We thank the reviewer for this helpful suggestion. We are very sorry for unclearly demonstrating our experimental results. The original comparison against an unadapted ChemDFM did not isolate the contribution of route-state information from that of local candidate features.

To address this concern, we introduce a matched Local-only ChemDFM control. It uses the same pretrained initialization, LoRA adaptation procedure, training examples, and listwise supervision as the full Analyzer. It retains all local candidate descriptors, including proposal rank, structural validity, and purchasable-reactant count, while excluding the unresolved frontier, route history, and remaining search budget.

Table R3-1: Contribution of route-state information in the Analyzer.

AnalyzerLocal featuresRoute-state featuresBeam Top-1Beam Top-3Beam Top-5Greedy DFS Top-1
Local-only ChemDFMYesNo36.6%46.3%49.8%35.52%
Full-state ChemDFMYesYes37.6%46.7%50.1%36.67%

The full-state Analyzer achieves consistent improvements across both beam search and greedy DFS, with gains of 1.0 and 1.15 percentage points in Top-1 accuracy, respectively. The improvements in Top-3 and Top-5 are smaller, suggesting that local candidate descriptors already provide a strong basis for candidate assessment.

These results indicate that route-state information serves as complementary planning context. By jointly considering candidate-specific properties and the evolving synthesis state, the full Analyzer achieves more effective branch evaluation across different search strategies. This supports the value of integrating local candidate assessment with state-aware decision making in RetroRoute.

Response to Question2:

We appreciate the opportunity to clarify the construction of the Analyzer's positive actions.

The Analyzer is trained using teacher-forced states derived from the annotated training synthesis routes. At each training state sts_{\text{t}}, the Reviewer generates a candidate set CtC_{\text{t}}. The positive subset is defined as

Pt=,ci∈Ct∣ci matches an annotated reference reaction action at st,.P_t=\\{\\,c_i\in C_t\mid c_i\text{ matches an annotated reference reaction action at }s_t\\,\\}.

Thus, positive labels are determined by reference-reaction matching at the corresponding training state. Neither URPO utility scores nor downstream rollout/search success is used to construct PtP_{\text{t}}.

The Analyzer is optimized using the multi-positive listwise objective:

=-\log\frac{\sum_{i\in P_t}\exp(z_i)} {\sum_{j\in C_t}\exp(z_j)}.$$ The two types of positive supervision serve different purposes. URPO uses complementary chemical and route-related utility signals to identify additional feasible alternatives for Reviewer refinement, whereas the Analyzer uses reference-route supervision to learn state-conditioned ranking among generated candidates. In particular, the Analyzer's positive labels are not inherited from URPO and do not represent observed downstream route-completion rewards. We hope this clarification resolves the reviewer's question about how $P_t$ is constructed and makes clear that the Analyzer's positive supervision is based solely on reference-reaction matching at teacher-forced training states, rather than on URPO utility scores or downstream search outcomes.
Reviewer_Pkn6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work uses LLMs to tackle multi-step retrosynthesis as a route-level decision-making problem. The low level of presentation and the serious omission of baselines make a favorable evaluation difficult.

Critical strengths
  • The paper takes a reasonable design approach to the modeling problem.
Critical weaknesses
  • Critical cherry-picking of baselines makes it impossible to judge the effectiveness or significance of the work itself. While these works are described as related work, the comparison is far from adequate. The authors use the MolT5 checkpoint created in RetroInText as the reviewer, yet there is no comparison with it at all. Since this omission appears to have been made knowingly, it will not be easy to reverse my evaluation even with additional baseline experiments during the rebuttal. This case is clearly different from one where baselines are thoroughly compared and the reviewer requests additional experiments for a paper the authors missed. Very important baselines such as RetroInText[1] and FusionRetro[2] are all missing.

[1] Kang, Chenglong, Xiaoyi Liu, and Fei Guo. "Retrointext: A multimodal large language model enhanced framework for retrosynthetic planning via in-context representation learning." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Liu, Songtao, et al. "FusionRetro: molecule representation fusion via in-context learning for retrosynthetic planning." International Conference on Machine Learning. PMLR, 2023.

  • URPO is the core contribution, but it is impossible to tell how it is defined.

  • The ablation study also lacks an analysis of URPO, experiments on why the LLM is necessary, and an analysis of the components of the planning scaffold.

  • The GitHub repository also does not appear to be anonymized.

Questions for authors

None.

8 comments
AuthorsOpenReview ↗

Thanks for the review - Weakness1(Part I)

Dear Reviewer Pkn6, We would like to express our sincere gratitude for your time and valuable feedback. Your constructive suggestions have been instrumental in refining our manuscript. We have carefully reviewed each comment and made the necessary revisions to address your concerns. Below, we provide our detailed point-to-point responses to the issues raised.

Response to Weakness1:

We thank the reviewer for raising concerns about the baseline coverage and the evaluation of RetroRoute. We acknowledge that the original manuscript did not include several important retrosynthesis methods in its main comparison. However, we respectfully disagree with the characterization of this omission as "critical cherry-picking" or the implication that stronger baselines were deliberately excluded to inflate our reported performance. An incomplete baseline comparison is a limitation that should be corrected, but it does not by itself establish intentional selection or suppression of competing results. We have taken this concern seriously and substantially expanded the comparison to provide a more complete and transparent assessment.

1. Expanded comparison with previously omitted baselines.

To address the concern about baseline selection, we provide a comprehensive summary of the published Retro* results reported in RetroInText[1], including template-based, semi-template-based, and template-free methods. This comparison explicitly includes RetroInText, NeuralSym, FusionRetro, and FusionRetro+CREBM, among other established baselines.

Table R2-1: Top-k exact-match accuracy (%) of retrosynthesis methods on RetroBench, with RetroRoute included for reference.

MethodTop-1Top-2Top-3Top-4Top-5
Template-based (Retro)*
RetroSim [2]35.140.542.944.044.6
NeuralSym [3]41.749.252.153.654.4
GLN [4]39.648.952.754.655.7
Semi-template-based (Retro)*
G2Gs [5]5.48.39.910.911.7
GraphRetro [6]15.319.521.021.922.4
GraphRetro+CREBM [7]16.320.121.622.322.7
Template-free (Retro)*
Transformer [8]31.340.444.747.248.9
Transformer+CREBM [7, 8]35.043.446.748.749.7
MEGAN [9]18.827.932.736.638.1
FusionRetro [10]37.545.048.350.651.5
FusionRetro+CREBM [7, 10]39.446.649.350.751.5
RetroInText [1]41.248.751.853.354.2
RetroRoute37.643.746.748.950.1

The expanded comparison makes the relative performance of these methods explicit. In particular, we acknowledge that RetroInText (41.2%), NeuralSym (41.7%), and FusionRetro+CREBM (39.4%) achieve higher published Top-1 accuracy than RetroRoute (37.6%) under their reported Retro* evaluations. We do not intend to obscure these results or claim that RetroRoute universally outperforms all previously published methods.

We believe that presenting these stronger baselines explicitly provides a more informative and fair assessment than restricting the comparison to a selected subset of methods.

2. Additional comparison under Greedy DFS.

Besides the published Retro* results, our manuscript also provides a comparison under Greedy DFS, which evaluates branch selection without relying on a complex beam-search procedure. Under this setting, only one local action is retained at each expansion, and the ground-truth route depth is used as the expansion limit following the corresponding RetroBench protocol.

Table R2-2: Greedy DFS Top-1 exact-match accuracy (%) on RetroBench.

MethodGreedy DFS Top-1
GraphRetro14.4
Transformer26.7
RetroSim31.5
MEGAN32.9
FusionRetro+CREBM33.8
FusionRetro35.2
RetroInText39.8
Neuralsym39.2
RetroRoute36.7

Under this evaluation, RetroRoute achieves 36.7% Top-1 accuracy, outperforming FusionRetro+CREBM and FusionRetro by 2.9 and 1.5 percentage points, respectively.

AuthorsOpenReview ↗

Thanks for the review - Weakness1(Part II)

3. Clarification of our methodological contribution and performance claims.

We would also like to emphasize that the contribution of RetroRoute is not simply the adoption of a pretrained single-step reaction predictor. Rather, it explores an alternative architecture for multistep retrosynthetic planning through the integration of three complementary components:

  • Planning-conditioned Reviewer: MolT5 generates reaction candidates conditioned on the target molecule and current planning context.
  • URPO: Utility-Regularized Positive Optimization refines the proposal distribution using complementary chemical feasibility and route-related utility signals.
  • State-aware Analyzer: ChemDFM evaluates candidate reactions under the evolving synthesis state, incorporating the unresolved frontier, route history, and remaining search budget to directly select branches for expansion.

Unlike conventional pipelines that separate single-step reaction generation from auxiliary pathway-ranking functions, RetroRoute employs the Analyzer as an active decision component in a closed-loop search process. Each selected reaction updates the planning state, which is then incorporated into subsequent decisions.

We have accordingly refined the positioning of our empirical claims to distinguish published cross-method benchmark results from comparisons under our evaluated search settings. Our contribution should be understood as an alternative LLM-driven architecture for sequential, state-aware retrosynthetic decision making.

We appreciate the reviewer for highlighting the importance of comprehensive baseline evaluation. We have addressed the substantive concern by explicitly incorporating the previously omitted methods, reporting their stronger published results where applicable, and providing additional comparisons under Greedy DFS. We believe these revisions provide a substantially more transparent and balanced evaluation of RetroRoute.

References

[1] Kang, C., Liu, X., and Guo, F. (2025). RetroInText: A Multimodal Large Language Model Enhanced Framework for Retrosynthetic Planning via In-Context Representation Learning. International Conference on Learning Representations (ICLR 2025).

[2] Coley, C. W., Rogers, L., Green, W. H., and Jensen, K. F. (2017). Computer-Assisted Retrosynthesis Based on Molecular Similarity. ACS Central Science, 3(12), 1237–1245.

[3] Segler, M. H. S., and Waller, M. P. (2017). Neural-Symbolic Machine Learning for Retrosynthesis and Reaction Prediction. Chemistry – A European Journal, 23(25), 5966–5971.

[4] Dai, H., Li, C., Coley, C. W., Dai, B., and Song, L. (2019). Retrosynthesis Prediction with Conditional Graph Logic Network. Advances in Neural Information Processing Systems (NeurIPS 2019), 32, 8870–8880.

[5] Shi, C., Xu, M., Guo, H., Zhang, M., and Tang, J. (2020). A Graph to Graphs Framework for Retrosynthesis Prediction. Proceedings of the 37th International Conference on Machine Learning (ICML 2020), 8818–8827.

[6] Somnath, V. R., Bunne, C., Coley, C. W., Krause, A., and Barzilay, R. (2021). Learning Graph Models for Retrosynthesis Prediction. Advances in Neural Information Processing Systems (NeurIPS 2021), 34, 9405–9415.

[7] Liu, S., Dai, H., Zhao, Y., and Liu, P. (2024). Preference Optimization for Molecule Synthesis with Conditional Residual Energy-based Models. Proceedings of the 41st International Conference on Machine Learning (ICML 2024), 30929–30945.

[8] Karpov, P., Godin, G., and Tetko, I. V. (2019). A Transformer Model for Retrosynthesis. International Conference on Artificial Neural Networks (ICANN 2019), 817–830.

[9] Sacha, M., Błaż, M., Byrski, P., Dąbrowski-Tumański, P., Chromiński, M., Loska, R., Włodarczyk-Pruszyński, P., and Jastrzębski, S. (2021). Molecule Edit Graph Attention Network: Modeling Chemical Reactions as Sequences of Graph Edits. Journal of Chemical Information and Modeling, 61(7), 3273–3284.

[10] Liu, S., Tu, Z., Xu, M., Zhang, Z., Lin, L., Ying, R., Tang, J., Zhao, P., and Wu, D. (2023). FusionRetro: Molecule Representation Fusion via In-Context Learning for Retrosynthetic Planning. Proceedings of the 40th International Conference on Machine Learning (ICML 2023), 22028–22041.

AuthorsOpenReview ↗

Thanks for the review - Weakness2

Response to Weakness 2:

We thank the reviewer for pointing out that the definition and operational procedure of URPO may not have been sufficiently clear. URPO has already been formally introduced in Section 3.3 (Eqs. (3)-(7)), with its implementation details provided in Appendix A.2. For clarity and ease of reference, we further elaborate on its formulation, candidate selection procedure, and policy optimization below.

URPO (Utility-Regularized Positive Optimization) is an offline, utility-guided positive policy refinement procedure. It aims to expand the proposal policy's support toward chemically feasible and route-supported alternative reactions.

Specifically, URPO starts from the Planning Scaffold SFT checkpoint. For each training context xx, the proposal policy generates up to 20 candidate reactions. The generated non-reference candidates are evaluated using multiple complementary utility signals, including molecular validity (uchemu_{\mathrm{chem}}), forward reaction consistency (ufwdu_{\mathrm{fwd}}), route compatibility (urouteu_{\mathrm{route}}), reference alignment (urefu_{\mathrm{ref}}), and building-block availability (ustocku_{\mathrm{stock}}), together with an undesirable-action penalty (PbadP_{\mathrm{bad}}). The aggregate utility is defined as:

U(c;x)={}&0.5u_{\mathrm{chem}}+1.0u_{\mathrm{fwd}}+0.5u_{\mathrm{route}}+0.3u_{\mathrm{ref}}+0.2u_{\mathrm{stock}}-1.0P_{\mathrm{bad}}. \end{aligned}$$ Here, $u_{\mathrm{chem}}$ assesses molecular validity; $u_{\mathrm{fwd}}$ evaluates whether an independent forward reaction model can reconstruct the current product; $u_{\mathrm{route}}$ measures compatibility with the synthesis trajectory available during training; $u_{\mathrm{ref}}$ measures structural and representational consistency with the demonstrated transformation; and $u_{\mathrm{stock}}$ rewards progress toward purchasable starting materials. $P_{\mathrm{bad}}$ penalizes undesirable reaction actions. Importantly, candidate selection is not determined solely by the aggregate utility score. As specified in Eqs. (4)-(5), URPO additionally imposes individual admissibility constraints to prevent a high score on one criterion from compensating for an unacceptable result on another. Specifically, the admissibility indicator is defined as: $$A(c;x)= \prod_{r=1}^{R} \mathbb{I}\left[q_r(c;x)\in I_r\right],$$ where $q_r(c;x)$ denotes the $r$-th candidate quality signal and $I_r$ its corresponding admissible region. The utility-supported positive action set is then constructed as: $$P_U(x) = \operatorname{Top} \left\\{ c \in C(x) \mid A(c;x)=1,\\; U(c;x)\geq\tau \right\\}.$$ In our implementation, the admissibility requirements include structural validity, zero undesirable-action penalty, forward plausibility of at least 0.5, and an aggregate utility score of at least 1.5 ($U(c;x)\geq 1.5$). Among the candidates satisfying these conditions, we retain the highest-utility actions, with at most one additional utility-supported action per original training example. Furthermore, the total number of additional positive examples is capped at 50% of the original gold training set size. These constraints preserve the demonstrated reaction distribution while allowing chemically and route-supported alternatives to receive additional policy support. The selected candidates are subsequently combined with the original demonstrated reactions to construct the augmented training set: $$D_{\mathrm{URPO}}=D_{\mathrm{gold}}\cup\left\\{(x,c)\mid c\in P_U(x)\right\\}.$$ The proposal policy is then further optimized using the standard positive log-likelihood objective: $$\begin{aligned} \mathcal{L}_{\mathrm{URPO}} = -\mathbb{E}_{(x,y)\sim D_{\mathrm{URPO}}} \log\pi_\theta(y\mid x). \end{aligned}$$ Therefore, URPO uses utility as an offline reward surrogate to identify and select additional positive training actions, rather than directly optimizing the policy through utility regression or online reinforcement learning. Rejected candidates are not required to serve as explicit negative examples. The policy update remains a standard supervised positive-likelihood objective, preserving training stability while incorporating additional utility-supported reactions. Finally, route-level supervision derived from synthesis annotations is used exclusively for offline candidate evaluation during training. These annotations are excluded from both the proposal policy input and the test-time search state. Overall, URPO provides a conservative mechanism for incorporating complementary chemical and route-related utility signals into proposal policy refinement, while retaining the stability of supervised learning. We hope this additional explanation clarifies both the mathematical definition and the practical implementation of URPO.
Reviewer_gzCpOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the mismatch between local reaction prediction and long-term route completion in multi-step retrosynthesis, as well as the ambiguity of treating non-reference reactions as negative examples. The proposed RetroRoute framework combines three components: (1) a planning-scaffolded Reviewer based on MolT5 that generates reaction candidates conditioned on the target, current intermediate, and search budget; (2) Utility-Regularized Positive Optimization (URPO), which augments supervised training with non-reference candidates selected using chemical validity, forward consistency, reference-route compatibility, and building-block availability; and (3) a state-aware Analyzer based on ChemDFM that performs listwise candidate reranking using the search frontier and route history.

Critical strengths

S1. Well-motivated separation of candidate expansion and state-aware decision making.

Decoupling positive-only candidate expansion from state-conditioned listwise reranking offers a reasonable framework for handling the multimodality of retrosynthetic reactions. URPO avoids explicitly treating all non-reference reactions as negative training examples, while the Analyzer evaluates competing candidates under a shared planning context. This separation is conceptually appealing and may improve candidate coverage without requiring a fully generative route-level decision policy.

S2. Consistent improvements in internal stage-wise ablations.

The experiments provide useful evidence that the proposed components improve performance within the authors' implementation. In Appendix B, single-step Top-1 accuracy increases from 44.1% with Planning Scaffold SFT to 47.8% after URPO refinement. In Appendix D, adding the adapted Analyzer improves Greedy DFS Top-1 accuracy from 31.98% to 36.67%. The depth-stratified results also show notable gains at route depths 4–6, suggesting that the combined framework improves reaction selection as planning decisions accumulate.

Critical weaknesses

W1. Incomplete baseline comparisons and inconsistent search protocols undermine the state-of-the-art claims.

The evaluation omits several stronger methods on the same RetroBench benchmark, including RetroInText, NeuralSym, and FusionRetro+CREBM. Notably, RetroInText, whose MolT5 checkpoint is used to initialize the Reviewer, reports 41.2% Top-1 accuracy under Retro*, compared with RetroRoute's reported 37.6%. Other omitted baselines also report higher accuracy under their published protocols.

More importantly, Table 1 presents the comparison as Retro* evaluation, whereas the released implementation uses a custom iterative beam search with a global depth limit of 14. Earlier RetroBench evaluations employ different search procedures and ground-truth-dependent depth limits. Without harmonizing these protocols, the reported performance differences cannot establish state-of-the-art superiority. This is a central concern because the claimed empirical advantage is one of the paper's primary contributions.

W2. The claimed learning of long-term route utility is insufficiently supported by the training objectives and ablations.

URPO selects additional positive reactions primarily through heuristic signals, including forward reaction consistency and overlap with reference-route molecules, rather than observed downstream route-completion outcomes. The released Analyzer training pipeline constructs teacher-forced reference states and assigns positive labels through single-step reference reaction matching, without implemented rollout-based supervision. Although this approach may indirectly capture useful planning information, the experiments do not establish that the model learns long-term action utility rather than reference-conditioned reaction preferences.

Furthermore, the Analyzer receives local candidate descriptors, including proposal rank, structural validity, and purchasable reactant count, alongside frontier and route-history information. The reported ablations do not isolate the contribution of these route-state inputs from that of local features and task-specific fine-tuning. The claimed mechanism of route-level decision making therefore remains insufficiently validated.

Questions for authors

Q1. Baseline Comparability and Evaluation Protocol

How does RetroRoute perform against stronger omitted baselines, particularly RetroInText, NeuralSym, and FusionRetro+CREBM, when evaluated using the same search algorithm, depth-limit protocol, and molecular matching criterion? Please also clarify which exact-match definition was used for the main results, given that the released evaluation code supports both full InChIKey and connectivity-only matching.

Q2. Evidence for Long-Term Route Utility

Given that URPO primarily relies on heuristic utility signals and the Analyzer is trained using teacher-forced reference matching without rollout-based supervision, what evidence demonstrates that the framework learns long-term route-completion utility rather than improving reference-conditioned single-step reaction selection?

Q3. Contribution of Route-State Information

How much of the Analyzer's improvement specifically comes from the unresolved frontier and route-history inputs? Can the authors provide a controlled comparison with an equally fine-tuned reranker using the same local candidate descriptors but excluding route-state context? Does the resulting policy generalize to different maximum search budgets?

Q4. Inference Efficiency

Does the complete RetroRoute framework retain its advantage over the URPO-only proposal model under matched inference-compute budgets, for example when the proposal-only system is allowed to explore a larger search beam using comparable computational resources?

10 comments
AuthorsOpenReview ↗

Thanks for the review - Weakness1 Part I

Dear Reviewer gzCp, We want to extend our heartfelt thanks for your time and give constructive suggestions. We have carefully considered each suggestion when revising this manuscript.Please see our point-to-point response to your concerns below.

Response to Weakness 1:

We sincerely thank the reviewer for identifying the incomplete baseline comparisons and the ambiguity in our evaluation protocol. We agree that the original manuscript did not sufficiently distinguish between published benchmark results and comparisons conducted under our own multistep search framework. We have therefore expanded the baseline comparison and revised the positioning of our contributions to provide a more accurate assessment.

1. Comprehensive comparison with existing RetroBench baselines.

We acknowledge that several important methods, particularly RetroInText, NeuralSym, and FusionRetro+CREBM, were not included in the original comparison. To address this omission, we provide a more comprehensive summary of the published Retro* results reported in RetroInText [1], covering template-based, semi-template-based, and template-free retrosynthesis approaches.

Table R1-1: Published Top-k exact-match accuracy (%) of retrosynthesis methods on RetroBench, with RetroRoute included for reference.

MethodTop-1Top-2Top-3Top-4Top-5
Template-based (Retro)*
RetroSim [2]35.140.542.944.044.6
NeuralSym [3]41.749.252.153.654.4
GLN [4]39.648.952.754.655.7
Semi-template-based (Retro)*
G2Gs [5]5.48.39.910.911.7
GraphRetro [6]15.319.521.021.922.4
GraphRetro+CREBM [7]16.320.121.622.322.7
Template-free (Retro)*
Transformer [8]31.340.444.747.248.9
Transformer+CREBM35.043.446.748.749.7
MEGAN [9]18.827.932.736.638.1
FusionRetro [10]37.545.048.350.651.5
FusionRetro+CREBM39.446.649.350.751.5
RetroInText [1]41.248.751.853.354.2
RetroRoute37.643.746.748.950.1

The expanded comparison provides a more representative picture of existing methods. We acknowledge that several published methods achieve higher accuracy than RetroRoute under their reported Retro* evaluations. Accordingly, we have refined our performance claims.

References

[1] Kang, C., Liu, X., and Guo, F. (2025). RetroInText: A Multimodal Large Language Model Enhanced Framework for Retrosynthetic Planning via In-Context Representation Learning. International Conference on Learning Representations (ICLR 2025).

[2] Coley, C. W., Rogers, L., Green, W. H., and Jensen, K. F. (2017). Computer-Assisted Retrosynthesis Based on Molecular Similarity. ACS Central Science, 3(12), 1237–1245.

[3] Segler, M. H. S., and Waller, M. P. (2017). Neural-Symbolic Machine Learning for Retrosynthesis and Reaction Prediction. Chemistry – A European Journal, 23(25), 5966–5971.

[4] Dai, H., Li, C., Coley, C. W., Dai, B., and Song, L. (2019). Retrosynthesis Prediction with Conditional Graph Logic Network. Advances in Neural Information Processing Systems (NeurIPS 2019), 32, 8870–8880.

[5] Shi, C., Xu, M., Guo, H., Zhang, M., and Tang, J. (2020). A Graph to Graphs Framework for Retrosynthesis Prediction. Proceedings of the 37th International Conference on Machine Learning (ICML 2020), 8818–8827.

[6] Somnath, V. R., Bunne, C., Coley, C. W., Krause, A., and Barzilay, R. (2021). Learning Graph Models for Retrosynthesis Prediction. Advances in Neural Information Processing Systems (NeurIPS 2021), 34, 9405–9415.

[7] Liu, S., Dai, H., Zhao, Y., and Liu, P. (2024). Preference Optimization for Molecule Synthesis with Conditional Residual Energy-based Models. Proceedings of the 41st International Conference on Machine Learning (ICML 2024), 30929–30945.

[8] Karpov, P., Godin, G., and Tetko, I. V. (2019). A Transformer Model for Retrosynthesis. International Conference on Artificial Neural Networks (ICANN 2019), 817–830.

[9] Sacha, M., Błaż, M., Byrski, P., Dąbrowski-Tumański, P., Chromiński, M., Loska, R., Włodarczyk-Pruszyński, P., and Jastrzębski, S. (2021). Molecule Edit Graph Attention Network: Modeling Chemical Reactions as Sequences of Graph Edits. Journal of Chemical Information and Modeling, 61(7), 3273–3284.

[10] Liu, S., Tu, Z., Xu, M., Zhang, Z., Lin, L., Ying, R., Tang, J., Zhao, P., and Wu, D. (2023). FusionRetro: Molecule Representation Fusion via In-Context Learning for Retrosynthetic Planning. Proceedings of the 40th International Conference on Machine Learning (ICML 2023), 22028–22041.

AuthorsOpenReview ↗

Thanks for the review - Weakness1 & 2 Part II

2. Clarification of search-protocol differences.

The published Retro* results should be distinguished from controlled comparisons using an identical search procedure. RetroRoute employs a native iterative beam-search controller with a route beam size of 5 and a global maximum search depth and number of search rounds of 14. Its ChemDFM-based Analyzer evaluates candidate reactions according to the evolving synthesis state, rather than relying on an additional A*-style path-value function.

For a more direct comparison, our manuscript also reports Greedy DFS results under the RetroBench depth-limit protocol. RetroRoute achieves 36.7% Top-1 accuracy, compared with 33.8% for FusionRetro+CREBM and 35.2% for FusionRetro. These results provide additional evidence of its effectiveness under a common greedy search setting.

3. Clarification of research objectives and methodological contributions.

RetroRoute aims to explore an alternative multistep retrosynthesis paradigm in which a chemistry-domain LLM actively participates in sequential route construction. Specifically, the MolT5-based Reviewer generates planning-conditioned reaction candidates, with URPO refining the proposal distribution, while the ChemDFM-based Analyzer directly evaluates candidate reactions under the evolving search state and selects branches for expansion.

Unlike conventional pipelines that separate reaction generation from auxiliary value-function-based route ranking, RetroRoute uses the Analyzer as a state-aware branch-selection component in a closed-loop search process. Its primary contribution therefore lies in integrating planning-conditioned reaction generation, utility-guided proposal refinement, and LLM-driven sequential branch selection, rather than claiming the highest published Top-k accuracy across all existing methods.

We hope these clarifications provide a more balanced comparison with existing baselines and better articulate the methodological contribution of RetroRoute.

Response to Weakness 2:

We sincerely thank the reviewer for this important observation. We agree that the original manuscript could have more clearly distinguished utility-guided proposal refinement, reference-supervised branch ranking, and long-horizon route optimization. In particular, our current training procedure does not directly optimize an observed downstream route-completion reward through rollout-based reinforcement learning. We therefore clarify the learning objectives and explain how route-level decision making is realized in RetroRoute.

1. The role of utility signals in URPO.

URPO is designed as an offline policy refinement mechanism. As defined in Section 3.3, it evaluates self-generated reaction candidates using complementary signals, including molecular validity, forward reaction consistency, route compatibility, reference alignment, and building-block availability.

We acknowledge that urouteu_{\mathrm{route}} and urefu_{\mathrm{ref}} incorporate synthesis annotations available during training. Specifically, urouteu_{\mathrm{route}} evaluates compatibility with the annotated synthesis trajectory, whereas urefu_{\mathrm{ref}} measures structural and representational consistency with the demonstrated transformation. These signals are therefore training-time utility proxies.

And URPO does not directly optimize a predicted long-term return. Instead, it uses the aggregate utility and individual admissibility constraints to identify additional positive reaction actions. These actions are combined with the original demonstrated reactions, and the proposal policy is subsequently refined using the standard positive log-likelihood objective. This design allows the proposal policy to incorporate complementary chemical and route-related criteria without requiring unreliable negative labels for unobserved reactions.

2. The distinction between Analyzer supervision and downstream route utility.

We also clarify that the Analyzer is trained using teacher-forced planning states derived from annotated synthesis routes. At each training state sts_t, the Reviewer generates a candidate set CtC_t, and the positive action subset is constructed according to reference-reaction matching:

Pt=,ci∈Ct∣ci matches an annotated reference reaction action at st,.P_t = \\{\\, c_i \in C_t \mid c_i \text{ matches an annotated reference reaction action at } s_t \\,\\}.

The Analyzer is optimized using the multi-positive listwise objective:

Llist=−log⁡∑i∈Ptexp⁡(zi)∑j∈Ctexp⁡(zj).\mathcal{L}_{\mathrm{list}}=-\log\frac{\sum_{i\in P_t}\exp(z_i)}{\sum_{j\in C_t}\exp(z_j)}.

Neither URPO utility scores nor downstream rollout/search success is used to construct these positive labels. Instead, the Analyzer learns to evaluate candidate reactions conditional on the current synthesis state. This distinction is important: URPO expands the proposal policy's support using utility-guided offline supervision, whereas the Analyzer learns state-conditioned preferences among generated candidate reactions using reference-route supervision.

AuthorsOpenReview ↗

Thanks for the review - Weakness2 Part III

3. How sequential route-level decision making is realized at inference.

Although the Analyzer is trained using teacher-forced states, its inference procedure operates within a closed-loop multistep search process. At each decision, it jointly considers the target molecule, current molecule, unresolved frontier, accumulated route history, remaining search budget, and the complete candidate set.

After an action is selected, the search controller expands the corresponding reaction and updates the unresolved frontier, route history, and current search depth. The subsequent Analyzer decision is therefore conditioned on the newly constructed partial route.

Furthermore, the route score accumulates state-conditioned decision probabilities across the complete action sequence:

\log p_{\phi}(a_t\mid s_t,C_t).$$ This mechanism couples individual reaction decisions through the evolving search state and allows earlier decisions to influence subsequent branch evaluation. We hope these clarifications make the actual learning mechanisms and the scope of our claims more precise.

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

Related papers

Discussion (0)

Sign in to comment.