acceptodds
Under review as a conference paper at ICLR 2027

Escalate Only When Necessary: Marginal-Value-Aware Feedback Acquisition for Retrosynthetic Planning

Abstract

Retrosynthetic planning decomposes a target molecule into building blocks, and recent LLM-based planners enrich this search with external feedback. Such feedback, however, varies substantially in fidelity and acquisition cost, creating a trade-off between decision reliability and resource consumption. We formulate feedback acquisition as adaptive escalation, where a planner progressively acquires higher-fidelity feedback only when its expected benefit justifies the additional cost. To learn such policies, we propose Marginal-Value-Aware Advantage Estimation (MAVE). MAVE models the reward-cost relationship across successive escalation levels, using its slope to estimate the current marginal value and its curvature to capture how this value evolves with further resource investment. MAVE incorporates both signals into advantage estimation, enabling marginal-value-aware credit assignment for just-enough escalation policy while discouraging premature stopping and over-escalation. Across three bencharks, MAVE achieves the highest success rates with hierarchical feedback, while reducing average feedback cost by 64.9% under the yield-only setting. Code is available at https://anonymous.4open.science/r/MAVE-retro/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Retrosynthesis, escalation, large language model

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

The paper studies adaptive feedback acquisition for retrosynthetic planning. MAVE fits a local reward-cost curve and incorporates its slope and curvature into GRPO advantages to learn when to stop or escalate feedback. Experiments evaluate planning performance and feedback cost across three benchmarks.

Critical strengths
  1. The problem is practically meaningful. Sequentially deciding whether stronger feedback is necessary provides a reasonable way to balance planning performance and acquisition cost, especially when feedback sources have different information and costs.

  2. The controlled comparisons and ablations are useful, the behavioral analyses help explain the learned policy. MAVE generally acquires deeper feedback for harder targets and preserves much of the benefit of fixed-depth acquisition at lower cost.

Critical weaknesses
  1. The reliability of the marginal-value estimates needs stronger support. The method fits a four-parameter tanh curve using only K=4K=4 stochastic rollouts per context, potentially with repeated costs. No fitting or derivative-stability diagnostics are reported. The ablations support the usefulness of the added terms, but do not establish that they reliably estimate the value of further feedback.

  2. The manuscript provides optimization hyperparameters but does not clearly specify the training/validation target sets, the concrete reward R(s,a)R(s,a), or the feedback budget BfB_f in Algorithm 2. These details affect generalization, reproduction, and the interpretation of the reward-cost trade-off.

Questions for authors

See Weaknesses.

Reviewer_3y6fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Marginal-Value-Aware Advantage Estimation (MAVE), an adaptive feedback escalation strategy for agentic multistep retrosynthetic planning. The method incorporates first- and second-order derivatives of the reward-cost relationship into advantage estimation to guide feedback acquisition. Across three benchmarks, MAVE achieves the highest planning success rates under a fixed budget of single-step model calls, showing improvements over existing methods.

Critical strengths
  1. Important problem formulation. Adaptive feedback acquisition is an important problem that reflects a promising direction for future agentic synthesis planning.
  2. Methodological novelty and empirical improvements. Incorporating first- and second-order derivatives into advantage estimation provides some methodological novelty. The method achieves the highest success rates across three benchmarks, demonstrating performance improvements.
Critical weaknesses
  1. Questionable formulation of feedback fidelity and cost. The costs and rewards associated with different external feedback levels are defined by the authors rather than grounded in actual tool execution costs or measured differences in feedback fidelity. This weakens the practical relevance of the proposed cost-aware escalation framework.
  2. Insufficient validation of synthesis quality and practical utility. The primary objective of hierarchical agentic synthesis planning should be to improve the quality of the resulting synthetic pathways. However, the experimental evidence relies heavily on planning success rates, which only indicate whether a path can be found within a hypothetical reaction network. This does not necessarily establish reaction feasibility or overall synthetic pathway quality. Furthermore, the experiments lack validation with genuinely high-cost, high-fidelity feedback, which may not be as reliable or informative in practice as assumed. Consequently, the experiments do not sufficiently support the paper's broader claims. The improvements over the strongest baselines are also relatively modest.
Questions for authors

The five feedback levels, ranging from no feedback to multistep synthesis strategies, appear to reflect a progression from local to global information rather than a hierarchy of actual acquisition costs or fidelity. What justifies this hierarchy, and how does it relate to the real computational costs and fidelity differences of the underlying tools?

Reviewer_4C6NOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a framework to select the feedback-level used by an LLM retrosynthetic planner, in an effort to mitigate the cost of feedback acquisition. This escalation policy is trained with RL, using an advantage estimate that includes the slope and curvature of a fitted reward-cost curve (MAVE). The authors compare their method to existing synthetic planners and show limited improvement in solve rate across 3 different datasets.

Critical strengths
  • The method proposed is simple and easy to follow.
  • The paper is well written, motivates the framework clearly, and supports it empirically.
Critical weaknesses

Two major concerns justify my current assessment. Below I describe them and propose ways the authors can convince me to increase my score:

  • [W1] The empirical results lack in performance and/or explanation: The paper shows limited performance gains by MAVE over the second-best baselines under the same hierarchical evaluation regime (Table 4), and route quality that degrades in Table 3 and does not improve consistently in Table 4. Below are some questions and suggestions to clarify the empirical impact of MAVE:

    • (a) The authors attribute the drop in route quality on 2 out of the 3 datasets in Table 3 to a quality-success trade-off (Section 5.2: routes might contain necessary low-yield intermediate steps which decrease the route quality). Can you support this explanation empirically (e.g. by reporting how often such steps are included in the routes found by MAVE), and discuss how much this drop matters in the assessment of the method's performance? Especially knowing that solve rate (the fraction of targets for which at least one route to purchasable building blocks is found) is a limited, heavily critiqued metric in the literature [1, 5].

    • (b) Can the authors add additional metrics to evaluate the quality of the routes found, independent of the feedback sources used during search (e.g. comparison to the reference routes of USPTO-190 and Wuxi-expert-873 through route similarity [2] or top-k route accuracy [3])? The current route quality is computed with the same yield predictor that provides the yield feedback (Appendix D.4), so the planner is evaluated by the model that guided it. Yield is also a single property of reactions that is notoriously difficult to learn [4].

    • (c) All methods are limited to 100 single-step model calls (Section 5.1), but this budget does not account other compute like the LLM calls of the escalation and reaction policies, the oracle calls, and the LLM-mediated baselines that may not use the single-step model at all. I suggest reporting the full computational resources used by each baseline, in the form of e.g. the number of LLM and oracle calls, the wall-clock time, the time to first solution, etc. Even better, the authors can constrain all methods with a common time budget following the recommendations of Syntheseus [5].

    • (d) How do non-LLM methods get yield as a feedback setting? They are not designed to take in any feedback by default, and it is therefore unclear which method the authors used to provide this feedback.

    • (e) Technically all the planners in Table 3 can take in hierarchical feedback, through modifying the value function in search [1, 6], filtering the precursors before expansion [7], or adding the feedback as input to the prompt of the LLM methods. I suggest the authors choose one of these ways to provide all their baselines with the same level of information, to isolate the effect of feedback availability vs MAVE's algorithm.

    • (f) The acquisition costs of the feedback sources are set by hand on a relative scale (Appendix A.1, Table 6), and the authors note that they do not reflect real-world expense. Since the paper is motivated by the cost of feedback, can the authors discuss how sensitive the results are to this cost schedule, e.g. by reporting results under an alternative schedule?

  • [W2] Neither the idea of choosing the feedback-level on a need basis nor how it is applied in this framework are novel, or particularly informative: adding feedback based on complexity and/or cost and favoring cheaper feedback first, seems like a clear design decision for this kind of system. The way the authors implement it is also straightforward. I am therefore struggling to see a significant scientific contribution in this paper. I invite the authors to do one of the following: 1) discuss why the idea of incorporating the first and second derivatives of a fitted reward-cost curve in the GRPO advantage estimate constitutes a meaningful contribution, 2) propose alternative ways to incorporate feedback in this setting and contrast it with their approach.

  • [W3] (optional) Minor modifications to the text presentation:

    • adding a table of contents at the beginning of the appendix would make navigating it easier.
    • Query cost and Query rate have no bolded/underlined values in the results tables.
Questions for authors

See weaknesses.

References

[1] Tripp, A., Maziarz, K., Lewis, S., Segler, M., and Hernández-Lobato, J. M. Retro-fallback: retrosynthetic planning in an uncertain world. ICLR 2024. https://arxiv.org/abs/2310.09270

[2] Genheden, S. and Shields, J. D. A simple similarity metric for comparing synthetic routes. Digital Discovery, 4:46–53, 2025. doi:10.1039/D4DD00292J.

[3] Genheden, S. and Bjerrum, E. PaRoutes: towards a framework for benchmarking retrosynthesis route predictions. Digital Discovery, 1:527–539, 2022. doi:10.1039/D2DD00015F.

[4] Voinarovska, V., Kabeshov, M., Dudenko, D., Genheden, S., and Tetko, I. V. When yield prediction does not yield prediction: an overview of the current challenges. Journal of Chemical Information and Modeling, 64(1):42–56, 2024. doi:10.1021/acs.jcim.3c01524.

[5] Maziarz, K., Tripp, A., Liu, G., Stanley, M., Xie, S., Gaiński, P., Seidl, P., and Segler, M. Re-evaluating retrosynthesis algorithms with Syntheseus. Faraday Discussions, 256:568–586, 2025. doi:10.1039/D4FD00093E.

[6] Yuan, L., Yu, Y., Wei, Y., Wang, Y., Wang, Z., and Wu, F. Active retrosynthetic planning aware of route quality. ICLR 2024. https://openreview.net/forum?id=h7DGnWGeos

[7] Segler, M. H. S., Preuss, M., and Waller, M. P. Planning chemical syntheses with deep neural networks and symbolic AI. Nature, 555:604–610, 2018. doi:10.1038/nature25978.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.