acceptodds
Under review as a conference paper at ICLR 2027

ChemOPD: Multi-Teacher On-Policy Distillation for Multi-Task Chemical Reasoning

Abstract

Large language models are increasingly expected to support diverse chemical reasoning capabilities within a unified model. One approach is to develop specialized capabilities separately and consolidate them through multi-teacher on-policy distillation, but this raises two questions: how should specialization be organized, and how should specialist guidance be integrated? We introduce ChemOPD, which addresses both. We estimate task affinities from supervised fine-tuning gradients and solve a constrained mixed-integer program(MIP) to construct partially overlapping specialist groups. During distillation, we retain a generalist teacher trained on all tasks so that specialist guidance supplements rather than replaces its supervision. Our anchor-residual objective gradually increases the routed specialist's contribution on student-generated responses. On ChemCoTBench, affinity-guided specialization produces task-dependent gains over the generalist teacher and improves several capabilities beyond semantic task grouping. Yet stronger teacher-side performance does not automatically yield stronger students: with the same specialists and routes, anchor-residual OPD improves most reported metrics over specialist-only distillation and realizes a larger share of the available teacher gains. These results highlight specialization and capability integration as connected but distinct design problems in chemical reasoning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Chemical Reasoning, AI4Science, Multi-Teacher On-Policy Distillation

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

This paper studies how to construct specialist teachers and integrate their capabilities into one chemical language model. It estimates task affinities from SFT gradients on disjoint training-data probes and uses a constrained mixed-integer program to form three partially overlapping task groups. Three Qwen3-8B specialists and an additional all-task Qwen3-8B generalist supervise a Qwen3-1.7B student on its own generated prefixes. The distillation advantage mixes generalist and routed-specialist signals, with the specialist weight increasing from 0 to 0.5. Evaluation covers 20 ChemCoTBench subtasks. Teacher comparisons, student baselines, component ablations, probe-resampling audits, and training diagnostics provide evidence of useful specialization and improved student performance, with task-dependent trade-offs.

Critical strengths
  1. The paper addresses a useful distinction between developing specialist capability and successfully integrating it. Importantly, ChemOPD and specialist-only MOPD share the specialist checkpoints, routes, student initialization, prompt stream, and actor-update budget. This makes their comparison informative about the supervision rule rather than merely stronger specialists.

  2. The teacher-construction stage and evaluation protocol are unusually well specified. The paper separates training memberships from distillation routes, reports teacher-side as well as student-side outcomes, and provides a probe-resampling audit that recovers the same grouping in ten draws.

Critical weaknesses
  1. Limited methodological novelty. The framework adds affinity-guided grouping and scheduled generalist-specialist mixing, but these are largely general-purpose extensions of multi-teacher OPD. The paper does not sufficiently establish what is fundamentally new about the combination of these existing components.

  2. Limited mechanistic understanding and generality. The results demonstrate useful performance gains, but provide limited explanation of why generalist-specialist integration improves capability transfer, or under what conditions it should be preferred over specialist-only distillation. The experiments are also restricted to Qwen3-8B teachers and a Qwen3-1.7B student, leaving generalization across different teacher-student capacity ratios unclear.

  3. Experimental robustness and ablation limitations. The warm-up ablation changes both the timing and average strength of specialist supervision (approximately 0.337 versus a constant 0.5), making it difficult to isolate the benefit of scheduling. An average-weight-matched constant baseline would help disentangle these effects. Additionally, a comparison with task-reward RL on suitable verifiable subtasks would help clarify the practical advantages over alternative post-training approaches.

Questions for authors

To what extent do the conclusions generalize beyond the Qwen3-8B to Qwen3-1.7B setup to other teacher-student sizes in chemistry? Please clarify which conclusions are experimentally supported and which remain hypotheses about broader multi-teacher distillation.

4 comments

Thank you for your careful review and suggestions on clarifying the methodological contribution, understanding capability transfer, and strengthening the capacity comparison and scheduling ablation.

Weakness 1 — Methodological contribution

We would like to clarify the methodological contribution of ChemOPD. Our novelty claim is not that gradient-based affinity estimation, constrained grouping, or linear interpolation is individually new. Rather, we address a broader and underexplored design problem in multi-teacher OPD for heterogeneous domains: the teacher collection itself must be constructed to reflect cross-task capability relationships before its knowledge can be effectively consolidated into a student.

This is particularly important for chemical reasoning, where benchmark task categories do not necessarily define independent training domains. Different tasks can share underlying chemical knowledge and benefit from complementary training signals. Treating the teacher collection as a fixed set of specialists trained on predefined, disjoint task categories may therefore overlook useful cross-task interactions. ChemOPD makes teacher construction an explicit part of the OPD framework. It uses estimated task affinities to determine specialist training memberships, permits selected tasks to contribute to multiple specialists, and separately maintains unambiguous prompt-level distillation routes.

The student integration stage addresses the complementary problem: even when specialists acquire useful capabilities, specialist-only OPD does not necessarily transfer these gains effectively. The generalist provides shared cross-task guidance, while the scheduled specialist contribution introduces task-specific corrections during on-policy learning. Thus, the two stages address different but connected requirements: constructing complementary sources of supervision and enabling a compact student to benefit from them. The controlled comparisons and the disjoint-grouping ablation provide empirical evidence for the respective contributions of these design choices.

We therefore position ChemOPD as a framework-level methodological contribution to multi-teacher OPD, rather than claiming a fundamentally new affinity estimator, MIP solver, or algebraically distinct interpolation operator.

Weakness 2 and Question — Transfer to a smaller student

We added the requested Qwen3-8B → Qwen3-0.6B comparison (Table R1 below). This increases the nominal teacher–student capacity ratio from approximately 4.7:1 to 13.3:1. We retain the existing teachers and task routes. At each student size, ChemOPD and MOPD share the SFT initializer, prompt stream, optimizer settings, and 109-update budget.

The overall advantage persists, while individual task outcomes change. ChemOPD improves addition P@1 at 0.6B (0.25 vs. 0.10), but MOPD leads at 1.7B (0.50 vs. 0.35). DRD2 gain shows the reverse pattern: ChemOPD leads at 1.7B but trails at 0.6B. Despite these changes, ChemOPD improves 22 of 31 metrics at both sizes, matching or improving 27/31 metrics at 0.6B and 26/31 at 1.7B (Table 1).

The logs show a similar change in training behavior at both sizes. Using the Appendix I diagnostics, ChemOPD reduces the extreme-candidate rate by 16.0% at 0.6B and 29.4% at 1.7B. Extreme-candidate rates and gradient norms before clipping are lower in every 10-step bin. The proportion of analyzed updates with gradient norm above 1.0 falls from 98.1% under MOPD to 83.5% under ChemOPD at 0.6B; at 1.7B it falls from 47.6% to 36.9%.

These results support the same central claim at both tested capacity ratios: generalist-anchored supervision moderates specialist-derived training signals and helps a compact student integrate heterogeneous capabilities, even though the gains differ by task and student size.

Table R1. Capacity comparison with a 0.6B student and 8B teachers.

The SFT student is the all-task initializer; the teacher column gives the routed 8B MIP-specialist results from Table 1. Bold marks ChemOPD results better than MOPD. Lower MAE is better; all other metrics are higher-is-better.

TaskMetricSFT student (0.6B)Routed teachers (8B)MOPD (0.6B)ChemOPD (0.6B)
AdditionP@10.250.500.100.25
DeletionP@10.400.700.500.50
SubstitutionP@10.020.620.170.22
Murcko scaffoldSimilarity0.430.770.440.47
EquivalenceAccuracy0.560.660.560.51
Functional-group countMAE0.120.080.090.08
Ring countMAE0.270.100.250.25
Ring-system scaffoldSimilarity0.620.820.400.68
QEDSR0.570.810.480.65
QEDDelta0.1210.1880.0980.158
LogPSR0.360.780.310.47
LogPDelta0.1020.7010.1800.295
SolubilitySR0.550.760.420.57
SolubilityDelta0.4500.7860.4120.609
DRD2SR0.360.590.220.37
DRD2Delta-0.0070.1360.0150.009
JNK3SR0.200.350.050.19
JNK3Delta-0.029-0.042-0.003-0.022
GSK3BSR0.300.580.110.22
GSK3BDelta0.0000.0660.005-0.014
Byproduct predictionFTS0.0180.0440.0230.028
Byproduct predictionTop-10.000.000.000.00
Major-product predictionFTS0.1520.4690.1750.181
Major-product predictionTop-10.010.260.000.00
Mechanism selectionAccuracy0.170.540.190.26
NEPPFTS0.2680.6400.2560.431
NEPPTop-10.010.350.000.04
Condition predictionFTS0.0080.0070.0000.015
Condition predictionTop-10.020.030.000.01
RetrosynthesisFTS0.2640.4820.2790.282
RetrosynthesisTop-10.000.030.000.00

Weakness 3 — Scheduling with a matched average weight

We added the requested control: the specialist weight stays at α = 36.75/109 ≈ 0.337156 throughout training, with anchor weight 1 − α. This matches the original schedule's average specialist weight. Teachers, routes, initialization, prompt stream, optimizer settings, and the 109-update budget are unchanged.

ChemOPD improves 23 of 31 metrics over this constant-weight baseline, with four ties and four losses. Substitution P@1 rises from 0.38 to 0.70, solubility gain from 0.468 to 0.718, and retrosynthesis FTS from 0.316 to 0.443 (Table 11; Table R2 below). The constant-0.5 control matches the final weight, while this new control matches the average weight. Together, they support a benefit from when specialist guidance is introduced, beyond simply lowering its average weight.

Comparison with task-reward RL. Thank you for this suggestion. Although some tasks have direct verifiers, their different output formats, lengths, and evaluation criteria require task-specific rewards and balancing. JNK3 and GSK3B also use pretrained random-forest oracles; using these predictors as rewards requires additional assessment of predictive uncertainty. This study focuses on capability consolidation through distillation. We plan a controlled task-reward RL comparison on verifiable subtasks in future work.

Table R2. Specialist-weight schedule ablation.

Constant 0.5 matches the final specialist weight, and constant 0.337 matches the average specialist weight. Both retain the generalist anchor. All variants use the original 1.7B student and 8B teachers. Bold marks ChemOPD improvements over constant 0.337. Lower MAE is better; all other metrics are higher-is-better.

TaskMetricConstant α=0.5Constant α≈0.337ChemOPD schedule
AdditionP@10.450.500.35
DeletionP@10.850.400.75
SubstitutionP@10.600.380.70
Murcko scaffoldSimilarity0.500.570.58
EquivalenceAccuracy0.570.520.62
Functional-group countMAE0.090.100.10
Ring countMAE0.200.150.15
Ring-system scaffoldSimilarity0.730.720.83
QEDSR0.840.820.86
QEDDelta0.1910.1860.167
LogPSR0.630.580.68
LogPDelta0.5000.5370.601
SolubilitySR0.610.600.71
SolubilityDelta0.4270.4680.718
DRD2SR0.450.500.55
DRD2Delta0.0550.0620.098
JNK3SR0.320.230.29
JNK3Delta-0.011-0.032-0.020
GSK3BSR0.440.480.46
GSK3BDelta0.0160.0220.029
Byproduct predictionFTS0.0340.0480.051
Byproduct predictionTop-10.000.000.00
Major-product predictionFTS0.3960.4820.465
Major-product predictionTop-10.150.140.18
Mechanism selectionAccuracy0.400.390.39
NEPPFTS0.5050.4900.521
NEPPTop-10.110.080.13
Condition predictionFTS0.0000.0000.023
Condition predictionTop-10.000.000.03
RetrosynthesisFTS0.3030.3160.443
RetrosynthesisTop-10.000.010.03

Thank you again for your valuable feedback. We would be glad to provide further clarification on any remaining questions about our submission.

Dear Reviewer Ugcp, as the rebuttal period is nearing its end, do you feel that our current rebuttal experiments and explanations have adequately addressed your concerns about our work? If so, would you consider updating your rating accordingly?

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

This paper proposes ChemOPD, a framework for training a unified, relatively small language model to handle diverse chemical reasoning tasks by combining the complementary capabilities of multiple specialized teachers. The paper addresses two connected but distinct challenges. First, how should heterogeneous chemical tasks be grouped when training specialist models? Second, once these specialists have been trained, how can their capabilities be effectively transferred into a single student without losing the benefits of general-purpose chemical reasoning?

Critical strengths
  1. A Clear Decomposition of the Multi-Task Learning Problem. The paper identifies two distinct challenges in consolidating chemical reasoning capabilities: organizing specialist expertise and transferring that expertise into a unified student. The affinity-guided grouping strategy provides a principled, gradient-based alternative to purely semantic task grouping, while the experiments suggest that specialization can yield task-dependent improvements. This decomposition is conceptually useful and provides a structured way to investigate the relationship between teacher construction and student capability acquisition.

  2. Controlled Comparisons Supporting the Value of Anchor-Residual Distillation. A notable strength is the comparison between ChemOPD and specialist-only on-policy distillation using the same specialists and routing assignments. The reported improvements on most metrics provide evidence that stronger specialists alone do not guarantee better student performance, and that retaining generalist supervision while gradually incorporating specialist guidance can improve capability transfer.

Critical weaknesses
  1. ChemOPD combines affinity-guided task grouping with an anchor-residual on-policy distillation strategy, but the paper does not sufficiently establish that the latter constitutes a fundamentally new or well-justified distillation mechanism beyond existing multi-teacher OPD approaches. Although the reported comparisons are encouraging, the methodological contribution would be more convincing with a clearer analysis of why the objective improves knowledge transfer and how its behavior differs theoretically or empirically from simpler teacher-combination strategies. 2.Insufficient experimental controls and robustness evidence. The reported improvements are evaluated on a single chemical benchmark, and some baseline comparisons do not fully match task exposure. The paper also provides limited evidence of variability across independent training runs, making it difficult to determine whether the reported gains are robust. In addition, the retrosynthesis evaluation uses a product-string-disjoint split rather than a canonical chemical-identity split, limiting the strength of the generalization claims.
Questions for authors

1.To what extent do the reported improvements stem from the anchor-residual distillation mechanism itself, rather than from teacher quality or task-routing strategies? 2.Can the authors demonstrate that the reported improvements persist under strictly matched task exposure and across multiple independent training seeds? In particular, could they report means and standard deviations for key metrics and clarify how potential data leakage from identical or highly similar chemical entities is addressed in the retrosynthesis evaluation, which uses a product-string-disjoint split?

2 comments

Thank you for your thoughtful assessment and constructive suggestions on the distillation mechanism, experimental controls, robustness, and retrosynthesis evaluation.

Weakness 1 and Question 1 — Contribution of the distillation rule

The main comparison holds teacher quality and routing fixed. ChemOPD and specialist-only MOPD use the same specialist checkpoints, task routes, SFT student initialization, prompt stream, and update budget (Section 5.1; Appendix C.1). This lets us test the supervision rule without changing the specialists or routing policy. ChemOPD records 22 wins, four ties, and five losses across 31 metrics. Deletion P@1 rises from 0.20 to 0.75, and solubility gain from 0.506 to 0.718 (Table 1; Appendix G, Table 11).

This comparison addresses the second of two problems introduced in Section 1: constructing useful specialists and transferring their capabilities to one student. MIP addresses the first. Its routed teacher profile improves 20 of 31 metrics over the generalist, including retrosynthesis FTS from 0.358 to 0.482 (Section 5.2, Figure 2). The human-defined controls show why intuitive task combinations are insufficient: Understanding+Reaction improves reaction performance while weakening understanding, and other combinations also introduce trade-offs (Appendix E, Table 10 and Figure 7).

The anchor helps the student use these stronger specialists. Equation (6) keeps the all-task generalist as a shared reference and gradually adds specialist guidance on the student's own prefixes. This is intended to moderate extreme specialist-derived corrections. Appendix G also compares the full method with a constant-0.5 mixture from the first update. ChemOPD matches or improves 24 of 31 metrics, including retrosynthesis FTS from 0.303 to 0.443. These comparisons support using the anchor and schedule together.

The training diagnostics help explain the observed benefit. ChemOPD reduces the extreme-candidate rate by 29.4% and the proportion of analyzed updates above the gradient-clipping threshold from 47.6% to 36.9% (Section 5.4, Figure 5; Appendix I). Median headroom recovery rises from 0.37 to 0.62 across 26 eligible metrics: the student retains more of the same teachers' improvement over SFT (Figure 4; Appendix H.1). These observations are consistent with more moderate training signals and better transfer under the controlled comparison.

Weakness 2 and Question 2 — Task exposure, robustness, and the retrosynthesis split

Matched task exposure. In the central ChemOPD–MOPD comparison, both methods share specialist checkpoints, routes, student initialization, shuffled prompt stream, and update budget (Section 5.1; Appendix C.1–C.2). The original runs both use 109 updates, 27,904 prompt draws, and 111,616 rollout sequences. Family exposure is also identical: 10,222 reaction, 12,138 editing/understanding, and 5,544 optimization prompts. Generated answers can differ because training is on-policy.

Homogeneous-RR and the DanceOPD-style controls match the total budget but change task exposure, as documented in Appendix C. Those comparisons test sampling strategies. The claim about the integration rule relies on the matched ChemOPD–MOPD comparison. We will make this distinction clearer.

Robustness across OPD seeds. We added two runs with different random seeds for each method, keeping the trained teachers, SFT initializer, routes, training data, and other hyperparameters fixed. Each paired comparison shares the prompt stream and update budget.

ChemOPD has better three-run means on 27 of 31 metrics, with one tie and three losses, and improves 17–22 metrics in each paired run. The gains cover editing, understanding, optimization, and reaction reasoning. Substitution P@1, ring-count MAE, QED success rate, and NEPP FTS improve in all three paired comparisons. This supports the robustness of the overall advantage across OPD seeds with fixed teacher and initialization checkpoints.

Three-run OPD results: 8B teachers to 1.7B student. Each cell reports the mean +- sample SD across three OPD runs (n = 3; the SD denominator is n - 1), using the task-level results in Table 1. Lower MAE is better; all other metrics are higher-is-better.

TaskMetricMOPD mean +- SDChemOPD mean +- SD
AdditionP@10.417 +- 0.0760.400 +- 0.087
DeletionP@10.400 +- 0.2000.733 +- 0.126
SubstitutionP@10.483 +- 0.0350.643 +- 0.060
Murcko scaffoldSimilarity0.487 +- 0.0310.540 +- 0.069
EquivalenceAccuracy0.547 +- 0.0600.577 +- 0.051
Functional-group countMAE0.107 +- 0.0310.093 +- 0.006
Ring countMAE0.217 +- 0.0290.150 +- 0.000
Ring-system scaffoldSimilarity0.757 +- 0.0320.797 +- 0.029
QEDSR0.763 +- 0.0060.847 +- 0.023
QEDDelta0.162 +- 0.0170.180 +- 0.016
LogPSR0.650 +- 0.0440.660 +- 0.020
LogPDelta0.550 +- 0.1020.591 +- 0.022
SolubilitySR0.660 +- 0.0260.693 +- 0.015
SolubilityDelta0.597 +- 0.1010.640 +- 0.072
DRD2SR0.493 +- 0.0120.467 +- 0.074
DRD2Delta0.058 +- 0.0270.073 +- 0.033
JNK3SR0.210 +- 0.0790.300 +- 0.017
JNK3Delta-0.032 +- 0.008-0.021 +- 0.003
GSK3BSR0.463 +- 0.0380.483 +- 0.025
GSK3BDelta0.032 +- 0.0110.035 +- 0.005
Byproduct predictionFTS0.034 +- 0.0030.047 +- 0.003
Byproduct predictionTop-10.000 +- 0.0000.000 +- 0.000
Major-product predictionFTS0.427 +- 0.0270.452 +- 0.013
Major-product predictionTop-10.143 +- 0.0310.147 +- 0.042
Mechanism selectionAccuracy0.413 +- 0.0320.400 +- 0.026
NEPPFTS0.455 +- 0.0460.528 +- 0.007
NEPPTop-10.110 +- 0.0200.127 +- 0.015
Condition predictionFTS0.021 +- 0.0070.055 +- 0.059
Condition predictionTop-10.027 +- 0.0060.040 +- 0.036
RetrosynthesisFTS0.431 +- 0.0060.439 +- 0.009
RetrosynthesisTop-10.013 +- 0.0060.023 +- 0.006

Chemical identity and similarity in the retrosynthesis split. We rebuilt the holdout because the released benchmark targets concatenate reactants without component separators (Section 5.1; Appendix B). We then checked chemical identity rather than relying only on different product strings. There are zero matches using atom-map-free canonical SMILES, with and without stereochemistry, and InChIKey. This holds against retrosynthesis training products and the structured molecular fields, including fragments, across all training tasks and affinity probes.

We also measured each test product's maximum Tanimoto similarity to any retrosynthesis training product using Morgan fingerprints (radius 2, 2,048 bits). The median is 0.367; 74% are below 0.5, only 8% reach 0.7, and none reach 0.8 (maximum 0.786). Most test products are therefore not close training analogues under this representation. For context, Klein et al., JCIM 2026 [1] describe ECFP4 scores below 0.5 as dissimilar and scores of at least 0.7 as very similar.

[1] Strategies for Identifying Molecules of Interest in Large Chemical Spaces

Evaluation on additional benchmarks. The current experiments use ChemCoTBench. We appreciate the suggestion and plan to test ChemOPD on additional chemistry benchmarks with heterogeneous tasks in future work.

We appreciate your encouraging assessment and helpful suggestions, which have helped us strengthen the experimental evidence and present our contributions more clearly.

Reviewer_uG6TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ChemOPD, a multi-teacher on-policy distillation framework for integrating heterogeneous chemical reasoning capabilities into a unified language model. The framework addresses two challenges: constructing specialist teachers through effective task grouping and transferring their capabilities into a shared student. To organize tasks, ChemOPD estimates task affinities using SFT gradients and formulates a constrained mixed-integer programming problem to identify partially overlapping task groups. Based on these groups, three specialist teachers and a generalist teacher are trained. For capability integration, the authors propose an anchor-residual distillation objective that combines generalist guidance with task-routed specialist supervision, gradually increasing the specialist contribution through a linear warm-up schedule. Experiments across 20 ChemCoTBench subtasks demonstrate task-dependent improvements from affinity-guided grouping, while the proposed distillation strategy outperforms specialist-only MOPD on most reported metrics under matched experimental conditions.

Overall, this paper presents a reasonably well-motivated framework for integrating specialized chemical reasoning capabilities through multi-teacher distillation. The combination of gradient-based task grouping and generalist-anchored specialist supervision provides a coherent approach to addressing potential conflicts between heterogeneous tasks. The controlled comparisons and additional ablations offer useful empirical support for the proposed design choices. Nevertheless, the improvements vary across tasks, and the evaluation remains concentrated on a single benchmark and a limited teacher configuration, leaving questions about the generalizability of the framework and the relative contributions of its individual components.

Critical strengths

1. Clear distinction between specialist construction and capability integration. The paper distinguishes improving teacher-side capabilities from transferring them to a unified student, supported by a controlled comparison with specialist-only MOPD under fixed specialist checkpoints, task routes, student initialization, prompt stream, and actor-update budget. The results show that stronger specialists do not necessarily yield stronger distilled students: specialist-only MOPD substantially degrades molecular deletion performance relative to the SFT initializer, while ChemOPD improves it. This motivates retaining generalist supervision in multi-teacher distillation.

2. Thorough experimental reporting and useful analyses. The paper details task grouping, training configurations, evaluation protocols, and the distillation objective, and contextualizes the approach via comparisons with semantic grouping, random grouping, and alternative teacher-sampling strategies. Component ablations, teacher-headroom analysis, and training-dynamics diagnostics further explain performance differences. Probe-resampling supports the stability of the selected task grouping, and the analysis includes both successful and unsuccessful chemical reasoning examples.

Critical weaknesses

1. Methodological contribution and overlapping-group benefits are not sufficiently isolated. This framework combines established gradient-based task affinity estimation, optimization-based task grouping, and multi-teacher on-policy distillation. Its advance over these approaches remains unclear. In particular, the anchor-residual objective in Eq. (6) is algebraically equivalent to a weighted interpolation between generalist and specialist advantages with a predetermined linear schedule. Experiments support this training strategy, but the paper does not distinguish it conceptually from scheduled mixing of teacher supervision signals. Overlapping task groups also deserve closer examination: in Table 8, only LogP and solubility are assigned to multiple specialist training groups, and the optimization specialist retains the same task composition as the semantic-grouping baseline. Comparisons with semantic and random grouping do not isolate the effect of allowing tasks to participate in multiple specialist groups. A comparison with disjoint affinity-based grouping and an ablation of the two shared tasks would provide direct evidence for the claimed benefit.

2. Robustness of downstream improvements and computational cost remain insufficiently characterized. Main training configurations use a single random seed and do not report variability across independent end-to-end teacher/student training runs. Ten probe-resampling experiments in Appendix F test the stability of affinity estimation and task grouping, not specialist or student training. This matters because overall comparisons summarize 31 heterogeneous metric columns, some with small between-method differences: ChemOPD improves retrosynthesis FTS from 0.437 to 0.443 over MOPD but performs worse on molecular addition. Without run-to-run variability, the reliability of smaller improvements and the consistency of the overall advantage are difficult to assess. In addition, ChemOPD requires both the generalist and the routed specialist to score student-generated prefixes, whereas specialist-only MOPD uses only the routed specialist. Although the methods share the same student update budget, their teacher-scoring costs are not equivalent. A quantitative comparison of training-time overhead would clarify the trade-off between the reported performance gains and the additional supervision cost.

Questions for authors

1. Contribution of overlapping task groups. Current comparisons with semantic and random grouping leave open whether the gains come from cross-group task sharing itself. A disjoint affinity-based grouping baseline would isolate this factor, and removing LogP and solubility from the reaction specialist’s training data would test the contribution of the two shared tasks directly. Where possible, such comparisons should control for training-data exposure or computational budget.

Would these additional experiments be feasible, and would they show that the observed improvements depend specifically on cross-group sharing?

2. Computational overhead of anchor-residual supervision. Anchor-residual supervision requires scoring student-generated prefixes with both the generalist and the routed specialist, so its teacher-scoring cost is higher than in specialist-only MOPD even when student update budgets match. Quantifying this overhead—teacher-scoring cost, total training GPU-hours, or wall-clock training time relative to specialist-only MOPD—would clarify the efficiency trade-off. How large is this additional cost in practice?

4 comments

Thank you for your careful review and helpful suggestions on clarifying the distillation design, isolating the contribution of task overlap, and assessing robustness and computational cost.

Weakness 1 and Question 1 — Distillation design and task overlap

Role of the generalist anchor. Equation (6) combines generalist and specialist advantages through scheduled interpolation. “Anchor-residual” describes their roles: the generalist provides guidance shared across tasks, and the routed specialist adds task-specific guidance. The aim is to help the student learn from specialists without receiving overly large corrections.

The controlled comparison with specialist-only MOPD keeps the specialists, routes, student initialization, prompt stream, and update budget fixed. ChemOPD improves most metrics under these conditions (Appendix G, Table 11). It also reduces the rate of extreme supervision signals by 29.4% and the proportion of analyzed updates above the gradient-clipping threshold from 47.6% to 36.9% (Figure 5; Appendix I). The student recovers more of the teachers' improvement over SFT: median headroom recovery rises from 0.37 to 0.62 (Figure 4; Appendix H.1). These results connect the two stages of ChemOPD: building stronger specialists and helping one student learn from them.

Grouping beyond chemical task labels. The MIP does more than share LogP and solubility. It considers individual subtasks without preset family boundaries, joins editing and understanding, and keeps the reaction subtasks together (Table 8). Agreement with some chemical categories shows that the affinity estimates recover sensible task relationships; the cross-category groups identify additional training relationships.

The human-defined grouping experiments explain why this matters. Combining Understanding+Reaction improves several reaction metrics but weakens understanding, and other combinations also introduce trade-offs (Appendix E, Table 10 and Figure 7). Output requirements, such as numerical answers versus molecular structures, may also affect which tasks train well together, although we have not isolated that factor. SFT-gradient affinities provide a measured basis for choosing groups when chemical labels alone do not identify the best training mixture.

Direct test of task overlap. We added the requested disjoint control. We set allowable overlap to zero by assigning each subtask to exactly one specialist, with the affinity matrix, teacher count, and group-size bounds unchanged. The MIP still operates on all 20 subtasks independently (Section 4.2; Appendix D). It selects Editing+Understanding, Optimization, and Reaction. This is exactly the original grouping with LogP and solubility removed from the reaction specialist, so the two requested controls coincide here.

Only the reaction specialist needs retraining. We retain the other two specialist checkpoints, generalist anchor, SFT student initializer, and OPD settings. The reaction examples still receive three SFT epochs, and the student receives the same 109 updates.

Without overlap, ChemOPD remains strong on editing/understanding: it records six wins, one tie, and one loss against MOPD across eight metrics. Across all tasks, it records 16 wins, four ties, and 11 losses (Table 1; table below). The unchanged editing/understanding specialist therefore remains useful in the disjoint setting.

The selected overlap provides a further benefit. Full ChemOPD improves 24 of 31 metrics over the disjoint variant, with four ties and three losses, including 11/12 optimization metrics and 9/11 reaction metrics. Solubility gain increases from 0.525 to 0.718, and retrosynthesis FTS from 0.395 to 0.443. The optimization teacher is unchanged, so the optimization differences occur in the jointly trained student. They are consistent with interactions between tasks during distillation.

Together with Section 5.2 and Appendix E, these results support using affinities to choose which tasks to combine, which to keep separate, and which to share across specialists for broader capability transfer.

Disjoint-grouping and shared-task ablation. Values follow Table 1. Bold marks no-overlap results better than MOPD. Lower MAE is better; all other metrics are higher-is-better.

TaskMetricMOPDChemOPD (no overlap)ChemOPD
AdditionP@10.500.550.35
DeletionP@10.200.800.75
SubstitutionP@10.520.520.70
Murcko scaffoldSimilarity0.480.580.58
EquivalenceAccuracy0.610.530.62
Functional-group countMAE0.140.110.10
Ring countMAE0.250.150.15
Ring-system scaffoldSimilarity0.770.780.83
QEDSR0.760.780.86
QEDDelta0.1670.1650.167
LogPSR0.700.670.68
LogPDelta0.6410.5990.601
SolubilitySR0.630.690.71
SolubilityDelta0.5060.5250.718
DRD2SR0.500.530.55
DRD2Delta0.0320.0230.098
JNK3SR0.180.270.29
JNK3Delta-0.032-0.025-0.020
GSK3BSR0.420.460.46
GSK3BDelta0.0200.0180.029
Byproduct predictionFTS0.0330.0480.051
Byproduct predictionTop-10.000.000.00
Major-product predictionFTS0.3980.3800.465
Major-product predictionTop-10.170.170.18
Mechanism selectionAccuracy0.450.380.39
NEPPFTS0.4620.4680.521
NEPPTop-10.130.080.13
Condition predictionFTS0.0280.0140.023
Condition predictionTop-10.030.040.03
RetrosynthesisFTS0.4370.3950.443
RetrosynthesisTop-10.010.010.03

Weakness 2 and Question 2 — Robustness and computational cost

Robustness across OPD seeds. We added two OPD runs with different random seeds for each method. The teachers, SFT initializer, routes, training data, and other hyperparameters remain fixed. Each ChemOPD–MOPD pair also shares its prompt stream and update budget.

Across the three runs, ChemOPD has better mean results on 27 of 31 metrics, with one tie and three losses. It improves 17–22 metrics in each paired comparison, so the overall advantage persists across the OPD seeds.

Retrosynthesis FTS varies little across these runs. ChemOPD reaches 0.439 ± 0.009, compared with 0.431 ± 0.006 for MOPD, and leads in two of the three paired comparisons. This supports an improvement on retrosynthesis, although the gain is smaller than on several other tasks.

Three-run OPD results: 8B teachers to 1.7B student. Each cell reports the mean +- sample SD across three OPD runs (n = 3; the SD denominator is n - 1), using the task-level results in Table 1. Lower MAE is better; all other metrics are higher-is-better.

TaskMetricMOPD mean +- SDChemOPD mean +- SD
AdditionP@10.417 +- 0.0760.400 +- 0.087
DeletionP@10.400 +- 0.2000.733 +- 0.126
SubstitutionP@10.483 +- 0.0350.643 +- 0.060
Murcko scaffoldSimilarity0.487 +- 0.0310.540 +- 0.069
EquivalenceAccuracy0.547 +- 0.0600.577 +- 0.051
Functional-group countMAE0.107 +- 0.0310.093 +- 0.006
Ring countMAE0.217 +- 0.0290.150 +- 0.000
Ring-system scaffoldSimilarity0.757 +- 0.0320.797 +- 0.029
QEDSR0.763 +- 0.0060.847 +- 0.023
QEDDelta0.162 +- 0.0170.180 +- 0.016
LogPSR0.650 +- 0.0440.660 +- 0.020
LogPDelta0.550 +- 0.1020.591 +- 0.022
SolubilitySR0.660 +- 0.0260.693 +- 0.015
SolubilityDelta0.597 +- 0.1010.640 +- 0.072
DRD2SR0.493 +- 0.0120.467 +- 0.074
DRD2Delta0.058 +- 0.0270.073 +- 0.033
JNK3SR0.210 +- 0.0790.300 +- 0.017
JNK3Delta-0.032 +- 0.008-0.021 +- 0.003
GSK3BSR0.463 +- 0.0380.483 +- 0.025
GSK3BDelta0.032 +- 0.0110.035 +- 0.005
Byproduct predictionFTS0.034 +- 0.0030.047 +- 0.003
Byproduct predictionTop-10.000 +- 0.0000.000 +- 0.000
Major-product predictionFTS0.427 +- 0.0270.452 +- 0.013
Major-product predictionTop-10.143 +- 0.0310.147 +- 0.042
Mechanism selectionAccuracy0.413 +- 0.0320.400 +- 0.026
NEPPFTS0.455 +- 0.0460.528 +- 0.007
NEPPTop-10.110 +- 0.0200.127 +- 0.015
Condition predictionFTS0.021 +- 0.0070.055 +- 0.059
Condition predictionTop-10.027 +- 0.0060.040 +- 0.036
RetrosynthesisFTS0.431 +- 0.0060.439 +- 0.009
RetrosynthesisTop-10.013 +- 0.0060.023 +- 0.006

Additional computational cost. We measured all methods on the same machine with eight GPUs and 109 updates. ChemOPD adds 15.3% to the OPD update cost relative to MOPD. In the original comparison, it improves 22 of 31 metrics, with four ties and five losses. We consider this a reasonable trade-off and will include the cost table and discussion in the revision.

MethodOPD update time (min)GPU-hoursRelative to MOPD
Vanilla OPD71.509.530.999×
Benchmark-family OPD78.2310.431.094×
MOPD71.549.541.000×
Homogeneous-RR67.529.000.944×
DanceOPD72.449.661.013×
Random-group OPD84.3211.241.179×
ChemOPD82.5011.001.153×

Thank you again for your constructive feedback. We would be happy to clarify any remaining questions about our submission.

Dear Reviewer uG6T, as the rebuttal period is nearing its end, do you feel that our current rebuttal experiments and explanations have adequately addressed your concerns about our work? If so, would you consider updating your rating accordingly?

Comments on the paper

1 comment

Dear Reviewers,

With the final hours of the discussion period approaching, we appreciate your engagement with our work. We're standing by if you have any additional questions.

Thank you for your thoughtful comments.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.