acceptodds
Under review as a conference paper at ICLR 2027

Beyond Action Entropy: Quotient-Space Exploration for Genome-Scale Metabolic Model Repair

Abstract

Repairing scientific models from functional observations differs fundamentally from supervised prediction: feedback may certify a solution without revealing which structural correction is responsible. We study this setting for genome-scale metabolic model (GEM) repair, where multiple reaction edits can explain the same phenotypes and many apparently distinct edits correspond to the same biological mechanism. This many-to-one structure creates a hidden failure mode for conventional exploration: diversity in the output space need not translate into diversity of scientific hypotheses. We introduce **QuotientPO**, which collapses equivalent repairs into canonical mechanisms and optimizes exploration directly over the resulting quotient space. To make quotient exploration informative under finite rollouts, we derive a kernelized R\'enyi estimator that resolves graded crowding among distinct repair cores beyond coarse exact-match counts. On 2,212 held-out GEMs, QuotientPO improves Success@32 from 17.93% to 20.10% (+12.1% relative) while consistently increasing distinct successful-core discovery under the same sampling budget. These results establish quotient-space exploration as a principled approach to mechanism-level discovery under verifier-induced equivalence.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Genome-scale Metabolic Models, Quotient Exploration, Reinforcement Learning, AI for Science

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xfrhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper repairs genome-scale metabolic models (GEMs) by adding and removing reactions so the model matches observed growth phenotypes. A flux balance analysis (FBA) verifier checks each repair.
  • The key idea is that many edit sets are equivalent under the verifier, so standard entropy bonuses reward redundant variants of the same repair. The authors instead reduce each repair to a "core" by greedily deleting edits that the verifier does not need, and treat repairs with the same core as one hypothesis.
  • The policy is a graph neural network that outputs a complete repair in one shot. It is trained in three stages: pretraining on GEM structure, SFT on solver-generated repairs, then RL. Their algorithm (QuotientPO) adds a diversity term to GRPO that discourages sampling cores similar to others in the rollout group.
Critical strengths
  • The paper is motivated by a clear conceptual insight. When many outputs are equivalent under the verifier, standard entropy bonuses reward representation multiplicity rather than real diversity. This failure mode has been studied before, both in RL literature and in LLM post-training. Applying it with verifier-induced equivalence classes in a scientific repair setting is natural and well motivated.
  • The authors assemble a large and diverse benchmark. They integrate public phenotype, genomic, and metabolic modeling resources into a unified repair task and a standardized verifier. They use this dataset both for training and evaluation.
  • The theory is correct as far as I checked, and the framing is clean.
  • They provide evidence that pretraining and SFT are needed before RL.
  • The experiments are broad. They compare against meaningful baselines, test robustness across RL objectives and policy architectures. They ablate each component to isolate the effects of individual parts.
Critical weaknesses
  • My main concern is the limited methodological novelty for an ML venue. The paper presents quotient-space exploration as a general principle, but most of its components already exist. The main new element is defining equivalence through the verifier via greedy deletion, but this is also the component with the weaknesses noted below.
  • Some closely related work is not discussed. Baram, Tennenholtz, Mannor (2021) show that action entropy in RL is biased by redundant actions, which is close to the paper's main motivation. Li, Zhang, Yu, Saha, et al. (2025) group LLM outputs into equivalence classes and rewards diversity across classes during RL, which is similar in spirit to quotienting before exploring.
  • The theory is mostly a combination of standard results. The framing is clean and appears correct, but the "theoretical foundation" label may be an oversell given the depth of the math.
  • The quotient space is defined per organism and per label set. The quotient is relative to whichever phenotypes were measured for that organism. As a result, the coarseness of the quotient space varies from organism to organism. For example, an organism tested on only two labels has few verifier-relevant edits, so many repairs collapse together. Since success-relateed metrics are macro-averaged across GEMs, it mixes quotients of very different granularity. This follows from the paper's design choice that identity is induced by the verifier, but it limits how much cores can be read as mechanisms.
  • The canonicalization procedure (Algorithm 1) deletes one edit at a time in a fixed order and stops when no single deletion preserves the label vector. Since the quotient space is defined through this specific procedure, it is hard to call it "canonical" in a strong sense. "Canonical" suggests an order-independent normal form, which this procedure does not provide. The math is unaffected, since the theorems only need a deterministic, verifier-preserving map. However, the interpretive claim that cores correspond to mechanisms relies on more than the algorithm guarantees.
Questions for authors
  • Did the authors consider simpler alternatives to RL techniques (RLOO/GRPO), such as rejection-sampling fine-tuning?
  • Can the deletion order be improved to get better results in the rest of the pipeline? Algorithm 1 sorts edits by (operation, reaction). Which operation is tried first? Sorting by operation first means all ADD (or all REMOVE) edits are tried for deletion before the other type, which introduces a systematic bias rather than just an arbitrary one. Did the authors experiment with other orderings? Would it be possible to order based on a biological prior?
  • How sensitive are the cores to deletion order? If the evaluation rollouts are canonicalized under several random fixed orders, how often does a sample's core change?
  • Is there a way to combine several valid verifier-preserving maps (κ1,…,κK\kappa_1, \dots, \kappa_K), for example the same greedy procedure under KK different fixed orders? Could such a combination result in a more robust core-reduction mapping? Could the map be alternated across rollout groups or updates, for example by drawing a different fixed deletion order each time? This would cost no more than a single map, and would reduce the risk that the diversity gains are tied to one particular partition.
  • QuotientPO is trained on the same partition that defines the core-based metrics. If the diversity gains reflect distinct mechanisms rather than artifacts of this particular partition, they should persist under an independently chosen map κ\kappa. Reporting such results would strengthen this claim.
  • SetPO (Li et al., 2026) also derives a leave-one-out marginal contribution from a kernelized set-diversity objective and uses it to shape advantages in policy optimization. Could the authors clarify which aspects of their crowding term are new relative to SetPO? Would it be possible to include SetPO as a baseline, adapted to the one-shot repair setting? If a direct comparison is not feasible, an explanation of why would also be helpful.
6 comments
AuthorsOpenReview ↗

Response to Reviewer xfrh (Part 1)

We thank the reviewer for recognizing our conceptual insight and comprehensive experiments.

W1-W3: Methodological Novelty Beyond Prior Work

We note that W2 refers to related works without complete bibliographic information (omitting titles and venues). We have made our best effort to identify the intended papers [1, 2] below. If these are not the works the reviewer intended, we would appreciate the reviewer's clarification.

1. Distinction from prior work. Our method addresses a fundamental challenge in scientific discovery: functional verification does not imply structural identification. A verifier can establish whether a candidate satisfies observed constraints without uniquely identifying the underlying mechanism. The central question is therefore what constitutes a distinct object of exploration, rather than simply how to encourage diversity. QuotientPO makes this question explicit and develops a unified methodology connecting representation, theory, and optimization:

  • Representation: From observable outcomes to distinct repair hypotheses. [1] characterize redundancy through MDP transitions, while [2] constructs learned semantic partitions. We go beyond these formulations by addressing a distinct identification problem: neither raw edit nor verifier-response provides an appropriate unit for exploring structural hypotheses under partial functional identification. We introduce an intermediate quotient through deterministic core reduction, removing verifier-redundant edits while retaining distinct irreducible repair cores as separate exploration targets. This establishes a functionally grounded representation between directly observable actions and verifier outcomes.

  • Objective: From diversity promotion to representation-invariant exploration. [2] promotes diversity through rewards over semantic partitions. We establish a more fundamental principle: exploration should depend on the distribution of repair hypotheses, not on how many raw representations realize each one. This makes representation invariance an explicit property of the exploration objective, rather than merely encouraging output diversity.

  • Optimization: From set diversity to principled quotient-space learning. [3] derives leave-one-out marginal diversity credit for advantage shaping. We instead derive group-relative crowding credit from a kernelized Rényi-2 entropy objective, yielding informative exploration signals even when exact-core collisions are sparse. Our formulation avoids intractable quotient marginalization while providing a direct connection between finite-sample updates and the population entropy gradient.

2. Theoretical foundation and empirical contributions.

  • Theoretical. We establish a theoretical foundation for representation-invariant exploration. Theorem 4.1 characterizes the necessary and sufficient condition for within-core invariance. For core-level utilities and unrestricted policies, we prove that action entropy induces representation-multiplicity bias, $q_{\mathrm{act}}^\star(m)\propto n_m e^{U_y(m)/\beta}$, whereas quotient entropy eliminates this bias, $q_{\mathrm{quot}}^\star(m)\propto e^{U_y(m)/\beta}$. We further establish finite-sample estimation guarantees and asymptotic local policy-gradient consistency for kernelized quotient exploration (Appendices B.3--B.4).

  • Empirical. We establish the first large-scale benchmark for phenotype-supervised amortized GEM repair (11,275 organisms, 2,212 held-out GEMs across five unseen genera). Under matched optimization settings, QuotientPO consistently outperforms baslines.

3. Our broader ML contribution is a problem-driven perspective on advancing ML through scientific challenges. Instead of treating scientific domains merely as downstream applications of existing algorithms, we use the partial identification inherent in graph-structured model repair to expose a fundamental limitation of conventional exploration. This illustrates how the structural properties of scientific problems can motivate new ML formulations and principles. QuotientPO provides a concrete example of this science-to-methodology direction, connecting AI4Sci with Sci4AI.

[1] Baram, Tennenholtz, and Mannor. Action Redundancy in Reinforcement Learning. UAI, 2021.

[2] T. Li et al. Jointly Reinforcing Diversity and Quality in Language Model Generations. arXiv:2509.02534, 2025.

[3] C. Li et al. SetPO: Set-Level Policy Optimization for Diversity-Preserving LLM Reasoning. ICML, 2026.

AuthorsOpenReview ↗

Response to Reviewer xfrh (Part 2)

W4: Quotient Granularity across Phenotype Label Counts

We thank the reviewer for this question.

Organism-specific quotients support consistent evaluation. Each GEM defines its own phenotype-conditioned quotient, and Successful Cores@32 counts distinct cores only within that GEM, without comparing core identities across organisms. Although quotient granularity may vary with phenotype coverage, Vanilla GRPO and QuotientPO are evaluated on the same GEMs under identical verifiers and reduction maps. Thus, both methods are evaluated against identical organism-specific criteria, ensuring a fair, paired comparison before macro-averaging.

Performance gains persist across phenotype-label counts. We stratify all 2,212 held-out GEMs by distinct measured phenotype labels and evaluate the original frozen 32-sample rollouts across 3 seeds. The table reports mean Success@32 (%) and successful cores per GEM. QuotientPO consistently outperforms Vanilla GRPO in both Success@32 and Successful Cores@32 across all strata, confirming that its gains do not depend on phenotype label density.

MetricMethod1–4 labels5 labels6–16 labelsAll
GEMs10324507302212
Success@32 (%)Vanilla GRPO22.7115.4812.6917.93
QuotientPO25.0717.9314.4320.10
Successful Cores@32Vanilla GRPO0.590.450.410.50
QuotientPO0.660.520.470.57

Fewer phenotype labels do not consistently yield greater core merging. We directly measure edit reduction and empirical core merging among distinct sampled repairs. Across the 1-4, 5, and 6-16 label strata, empirical merge rates are 56.46%, 61.10%, and 56.35% for Vanilla GRPO, and 55.32%, 58.25%, and 53.30% for QuotientPO. Corresponding edit reduction rates are 94.14%, 94.79%, and 93.65% for Vanilla GRPO, and 94.94%, 94.97%, and 93.83% for QuotientPO. Empirical merging peaks in the 5-label stratum rather than the lowest-label stratum, indicating that fewer measured phenotypes do not systematically increase core merging.

To further establish the biochemical significance of core diversity, we identify distinct repair interventions using reaction stoichiometry, compartment information, and FBA counterfactuals (see our response to Reviewer hPQr, W2&Q1 and Appendix G).

W5: Canonicalization and Biological Mechanisms

We thank the reviewer for raising this concern.

Verifier-preserving core reduction is theoretically sound and empirically robust. We use canonicalization to denote a fixed-order, verifier-preserving core reduction, rather than an order-independent biological normal form. It satisfies Theorem 4.1 without requiring globally minimal representatives. Across four deletion orders, QuotientPO retains a successful-core coverage advantage of +0.0558 to +0.0685 cores per GEM. Only 14 of 29,876 distinct starting cores admit further pairwise reduction, with virtually unchanged gains under stronger reduction (+0.0685 vs. +0.0684). (see also our response to Reviewer KTZQ, W2)

Biologically distinct mechanisms are supported by independent functional evidence. Beyond core identity, we identify distinct biochemical interventions through reaction stoichiometry, compartments, and FBA focal-reversion tests. QuotientPO improves their joint recovery, notably for alternative nitrogen donors in ASNS1 (33.79% vs. 14.70% DualRecovery@32).(see also our response to Reviewer hPQr, W2&Q1) These findings support the discovery of biochemically distinct repair hypotheses without equating each core with a unique biological mechanism.

Q1: Rejection-Sampling Fine-Tuning

We thank the reviewer for suggesting this comparison. We evaluate two online RFT baselines initialized from the same Soft-SFT checkpoint: RFT-Success retains phenotype-consistent repairs, while RFT-Improved retains repairs with improved phenotype consistency. Both use Bernoulli NLL with GEM-balanced replay under matched architectures, verifiers, and training budgets across 3 seeds. As shown below, both RFT variants improve Success@32 over Soft-SFT with limited gains at Success@1, but remain below Vanilla GRPO and QuotientPO across all metrics. These results highlight the advantage of policy-gradient optimization over verifier-filtered self-training, with further gains from quotient-space exploration.

MethodSuccess@1 (%)Success@32 (%)Successful Cores@32
Soft-SFT2.40±0.052.40 \pm 0.057.52±0.497.52 \pm 0.490.16±0.010.16 \pm 0.01
RFT-Success2.70±0.212.70 \pm 0.2113.20±0.8413.20 \pm 0.840.287±0.0170.287 \pm 0.017
RFT-Improved2.38±0.152.38 \pm 0.1512.73±0.2112.73 \pm 0.210.257±0.0040.257 \pm 0.004
Vanilla GRPO7.43±0.037.43 \pm 0.0317.93±1.5417.93 \pm 1.540.50±0.010.50 \pm 0.01
QuotientPO7.76±0.23\mathbf{7.76 \pm 0.23}20.10±0.52\mathbf{20.10 \pm 0.52}0.57±0.02\mathbf{0.57 \pm 0.02}
AuthorsOpenReview ↗

Response to Reviewer xfrh (Part 3)

Q2: Reduction-Order Selection and Biological Priors

Algorithm 1 uses ADD-first ordering with reaction-ID tie-breaking, affecting core selection but not verifier preservation. We evaluate 6 reduction orders on identical frozen test rollouts (2,212 GEMs, 32 samples each,3 seeds), including an organism-specific biological prior. Reaction-level support is derived from DIAMOND homologs (≥30%\geq30\% identity, ≥70%\geq70\% bilateral coverage, E-value ≤10−5\leq10^{-5}) and aggregated through gene-protein-reaction (GPR) rules, prioritizing the retention of protein-supported reactions. All reductions preserve verifier responses and single-edit irreducibility. QuotientPO consistently outperforms Vanilla GRPO in successful-core coverage across all orders, while the biological prior also outperforms its shuffled control for both methods.

Reduction orderVanillaQuotientPOPaired Δ\Delta
ADD-first0.5016±0.01290.5016 \pm 0.01290.5701±0.01980.5701 \pm 0.0198+0.0685±0.0168+0.0685 \pm 0.0168
REMOVE-first0.5839±0.04560.5839 \pm 0.04560.6510±0.01200.6510 \pm 0.0120+0.0671±0.0476+0.0671 \pm 0.0476
Random-fixed 10.5577±0.02470.5577 \pm 0.02470.6135±0.00550.6135 \pm 0.0055+0.0558±0.0212+0.0558 \pm 0.0212
Random-fixed 20.5824±0.02830.5824 \pm 0.02830.6444±0.01830.6444 \pm 0.0183+0.0619±0.0103+0.0619 \pm 0.0103

Reduction orderVanillaQuotientPOPaired Δ\Delta
Shuffled prior0.5270±0.03340.5270 \pm 0.03340.5761±0.02050.5761 \pm 0.0205+0.0491±0.0219+0.0491 \pm 0.0219
Biological prior0.5631±0.03880.5631 \pm 0.03880.6260±0.00210.6260 \pm 0.0021+0.0628±0.0367+0.0628 \pm 0.0367

Q3: Sensitivity of Core Identity to Reduction Order

We quantify core-identity sensitivity by re-reducing identical frozen successful repairs under three alternative fixed orders. Relative to ADD-first, the fractions of successful repairs whose core identities change are 33.52%/35.83% (REMOVE-first), 50.67%/52.74% (Random-fixed 1), and 55.97%/57.93% (Random-fixed 2) for Vanilla/QuotientPO. These changes reflect alternative single-deletion-irreducible representatives preserving identical verifier responses. Even globally minimum-cardinality repairs need not be unique. Crucially, Theorem 4.1 applies to any fixed verifier-preserving quotient, without requiring order-independent core identities. Although deletion order affects the resulting partition, both methods exhibit comparable sensitivity, and QuotientPO maintains higher successful-core coverage across all tested mappings (Q2). Thus, the exploration advantage persists despite substantial variation in core identity.

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

The paper propose a method for reparing a genome-scale metabolic model (GEM). A GEM is a bipartite graph over metabolites and reactions. The proposed method introduces equivalent classes to get more relevant diversity when doing the repairs. The repairs are proposed reaction additions/deletions.

Critical strengths

The paper propose an interesting way of imposing meaningful diversity in the GEM repair task.

Improves on all metrics compared to all baselines. The paper also includes relevant ablations.

Critical weaknesses

Would be nice with a basic explanation of the setup (what is the problem, what data do you have at train vs test time, etc.). It was difficult to understand it. Figure 1 wasn't too helpful either. Better communication would help making the advantages of the method more clear.

I'm not sure how a user selects among candidate repairs when using the model on a dataset without growth data.

Questions for authors

When using the model without growth observations, is the intended output a single selected repair or a set of hypotheses?

Could you explain the setup in basic terms?

6 comments
AuthorsOpenReview ↗

Response to Reviewer hPQr (Part 1/2)

We thank the reviewer for recognizing our meaningful-diversity formulation and empirical results. We clarify the scientific setting and the practical use of repair hypotheses.

W1 & Q2: Scientific Motivation and Problem Formulation

Task definition: repairing structural errors in draft metabolic networks using genome and network context. A genome-scale metabolic model (GEM) represents an organism's metabolism as a biochemical reaction network. Draft GEMs compiled from genomes by standard bioinformatics pipelines (e.g., CarveMe) frequently contain missing or spurious reactions, leading to discrepancies between FBA-predicted growth and experimentally observed phenotypes. Our task is to predict discrete reaction additions and deletions that restore phenotype-consistent network functionality.

Training and Inference: phenotype-supervised training enables phenotype-free deployment. Metagenomics enables genome reconstruction without cultivation, while many microorganisms remain difficult to culture, limiting access to experimentally measured growth phenotypes [1]. Unlike classical phenotype-guided GEM refinement [2], we learn transferable reaction additions and deletions across organisms using FBA-based rewards and offline teacher generation, without gold-standard corrected GEMs. This design transfers phenotype supervision across organisms, enabling GEM repair even when growth measurements are unavailable at deployment.

Necessity of quotient exploration: quotienting out hitchhiking edits prevents spurious diversity and drives discovery of distinct repair mechanisms. Because FBA only checks phenotype satisfaction, adding irrelevant reactions to a valid repair yields the exact same reward. Standard action-space exploration treats such redundant variants as distinct solutions, wasting exploration on the same underlying fix mechanism. QuotientPO prunes each solution down to its essential functional core, forcing the policy to explore fundamentally different metabolic pathways rather than cosmetic variations.

[1] Nayfach et al. A genomic catalog of Earth's microbiomes, Nature Biotechnology, 2021.

[2] Arkin et al. KBase: The United States Department of Energy Systems Biology Knowledgebase. Nature Biotechnology, 2018

AuthorsOpenReview ↗

Response to Reviewer hPQr (Part 2/2)

W2 & Q1: Deployment and Scientific Value of Diverse Repairs

We thank the reviewer for this practical question. Without test-time growth observations, our trained policy supports both predictive deployment (outputting a single repair or a prioritized shortlist) and scientific exploration (generating a diverse pool of candidate hypotheses).

For predictive deployment, sampled repairs can be prioritized without test-time growth observations. When a single corrected GEM is needed, greedy decoding directly outputs one deterministic repair, whereas stochastic sampling generates a 32-candidate pool. To select among sampled candidates without test-time phenotypes, Policy ranks repairs by their model log-likelihood, while Diversity starts from the highest-likelihood repair and greedily maximizes the minimum signed-edit Jaccard distance to previously selected candidates. Evaluated across the test set (32 candidates per GEM), QuotientPO consistently improves shortlist success rates across all budgets under both strategies:

Method@1@2@4@8
Vanilla + Policy7.939.7811.2113.20
QuotientPO + Policy8.2110.4312.5414.69
Vanilla + Diversity7.9310.1612.2114.35
QuotientPO + Diversity8.2110.6213.2215.57

For scientific discovery, diverse repair hypotheses enable the recovery of distinct biochemical interventions. Beyond our qualitative case studies on nitrate utilization, proline supply, UTP synthesis, and asparagine synthesis (Fig. 8 and Appendix G), we quantitatively evaluate across the test set whether increased core diversity translates into discovering distinct biochemical routes for the same phenotype constraints. We use held-out phenotype constraints, reaction stoichiometry, compartments, and FBA counterfactuals to verify biochemically distinct interventions recovered in different successful repair cores. Across four representative biochemical families, DualRecovery@32 (metric used in the table below) measures how often 32 sampled repairs recover both alternatives, evaluated on a shared cohort where either method recovers at least one verified route. QuotientPO achieves higher joint recovery across all four families, demonstrating that its diversity gains extend beyond distinct core counts to biochemically meaningful alternative repairs:

Biochemical AlternativeTargetVanilla (%)QuotientPO (%)
Alternative sodium transport routesna136.3838.49
Nitrite transport vs. intracellular productionno24.505.58
Asparagine synthesis: glutamine vs. ammoniumASNS114.7033.79
Ornithine-to-putrescine conversion across compartmentsHMR_442228.5232.96

Revisions. We have revised the caption of Figure 1 and the first paragraph of Introduction section to make the training/deployment distinction explicit, and added phenotype-free shortlist evaluations in Appendix F.8 and quantitative analysis of biochemical alternative recovery in Appendix G.2.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer hPQr,

Thank you again for your thoughtful review and constructive feedback. We have provided detailed clarifications and additional experiments addressing your concerns.

We would greatly appreciate any further feedback or remaining concerns you may have after reviewing our responses. We would be happy to provide additional clarification if needed.

Best regards,

Paper 20414 Authors

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

This paper studies phenotype-supervised genome-scale metabolic model (GEM) repair, aiming to predict reaction additions and deletions without requiring experimental phenotype labels at test time. The authors propose QuotientPO, a reinforcement learning method that encourages diversity among functionally distinct repair cores rather than raw reaction-edit sets. Specifically, the method uses verifier-preserving canonicalization to remove redundant edits and introduces kernelized Rényi entropy to measure diversity among canonical repair cores. The resulting diversity objective is integrated into GRPO to improve exploration. Experiments demonstrate that QuotientPO outperforms existing baselines in both repair success rate and distinct successful-core discovery.

Critical strengths
  1. Application-motivated method design. The proposed method is well motivated by the characteristics of GEM repair, where different reaction-edit sets may represent redundant repair mechanisms. By considering diversity at the level of canonical repair cores rather than raw edits, QuotientPO provides an application-aware exploration strategy that aligns well with the structure of the problem.
  2. Consistent performance improvements over existing baselines. QuotientPO achieves the best performance among the evaluated learning-based methods. It also increases distinct successful-core discovery, demonstrating improvements in both repair success and solution diversity.
Critical weaknesses
  1. Limited modeling of dependencies between reaction edits. The proposed method models each legal reaction edit using an independent Bernoulli distribution. However, reaction edits may have interdependent effects, where multiple edits must be applied jointly to achieve a successful repair. The independence assumption may therefore limit the policy's ability to capture such dependencies and represent diverse valid repair combinations.
  2. Limitations of greedy deletion-based canonicalization. The proposed canonicalization greedily removes individual edits while preserving verifier outcomes. However, this procedure does not account for dependencies among edits: multiple edits may be jointly redundant even when removing any single edit changes the verifier outcome. This limitation may lead to inconsistent representations of functionally equivalent repairs, potentially affecting the reliability of the quotient-space diversity objective.
Questions for authors

See Weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer KTZQ (Part 1/2)

We thank the reviewer for recognizing our method's motivation and consistent empirical improvements. We address the concerns (W1-W2) below.

W1: Limited modeling of dependencies between reaction edits.

We appreciate the opportunity to clarify this design choice. Conditional independence during action sampling does not imply functional independence in metabolic effects, nor does it prevent learning coordinated multi-edit repairs. We clarify this from three key points:

1. Global context conditioning and joint verification capture interacting metabolic effects. Although edits are sampled independently given the input, their probabilities are predicted by a shared GNN encoding the full metabolic network. More importantly, FBA evaluates the entire sampled repair set jointly, providing a set-level reward that directly reinforces coupled multi-reaction interactions during policy-gradient updates (Section 3, Appendix A.1).

2. Factorization does not prevent representing coordinated multi-edit repairs. For any target repair E⋆E^\star, the factorized Bernoulli policy assigns over candidate edits aa probability: π(E⋆∣x)=∏a∈E⋆pa(x)∏a∉E⋆(1−pa(x)),\pi(E^\star\mid x) = \prod_{a\in E^\star}p_a(x)\prod_{a\notin E^\star}\bigl(1-p_a(x)\bigr), which satisfies π(E⋆∣x)→1\pi(E^\star\mid x) \to 1 as pa(x)→I[a∈E⋆]p_a(x) \to \mathbb{I}[a \in E^\star] for all aa. Thus, factorization does not limit representing coupled repairs, even when individual edits are ineffective in isolation. In practice, QuotientPO readily discovers such coordinated edits, such as the joint removal of NITR and NTR3B (Figure 8(a)).

3. Quotient exploration improves diverse repair discovery across policy architectures. Quotient-space exploration operates on complete, jointly verified repair cores and does not require an explicitly correlated action decoder. Under the identical Soft-policy architecture, QuotientPO improves both repair success and successful-core coverage over Vanilla GRPO (Table 1). The improvement also persists under the alternative Hard-policy architecture (Table 3(b)). These comparisons demonstrate that QuotientPO consistently improves repair discovery across different policy architectures, without relying on a particular policy parameterization.

AuthorsOpenReview ↗

Response to Reviewer KTZQ (Part 2/2)

W2: Limitations of greedy deletion-based canonicalization.

We thank the reviewer for raising this concern. We show that higher-order redundancy does not compromise the theoretical validity of our approach, while establishing the computational advantages and empirical robustness of our reduction procedure below:

Theoretical validity of verifier-preserving reduction. Under a fixed deletion order, Algorithm 1 defines a deterministic, verifier-preserving mapping κy\kappa_y whose output is single-deletion-irreducible: σy(κy(E))=σy(E)\sigma_y(\kappa_y(E))=\sigma_y(E). While residual higher-order redundancy and alternative deletion paths may assign functionally equivalent repairs to distinct cores, κy\kappa_y never merges repairs with different verifier responses. Thus, the induced quotient is a well-defined refinement of verifier-response equivalence (Appendix A.2). Crucially, Theorem 4.1 applies to a fixed quotient without requiring globally minimal representatives.

Computational tractability without global minimization. Finding a minimum-cardinality verifier-preserving subrepair is NP-hard by reduction from Minimum Set Cover (Appendix A.2). Our deterministic reduction instead guarantees verifier preservation and single-deletion irreducibility with polynomially many verifier evaluations, avoiding combinatorial minimization during RL rollouts.

Empirical robustness to higher-order redundancy and deletion order. We conduct two complementary evaluations using identical frozen test rollouts from Vanilla GRPO and QuotientPO in our main results, without retraining or resampling.

  • (a) Higher-order redundancy is rare and has negligible impact on core coverage. We extend the original reduction to iteratively remove verifier-preserving singletons and edit pairs until neither can be removed. Among 29,876 distinct starting cores, only 14 (less than 0.1%) contain additional pairwise redundancy. Importantly, QuotientPO's paired Success Cores@32 advantage remains nearly unchanged (+0.0685 vs. +0.0684) under the stronger reduction.
ReductionVanillaQuotientPOPaired Δ\Delta
Original κ1\kappa_10.5016±0.01290.5016 \pm 0.01290.5701±0.01980.5701 \pm 0.0198+0.0685±0.0168+0.0685 \pm 0.0168
Enhanced κ2\kappa_20.5016±0.01290.5016 \pm 0.01290.5700±0.01960.5700 \pm 0.0196+0.0684±0.0170+0.0684 \pm 0.0170
  • (b) Coverage gains persist across alternative deletion orders. Re-canonicalizing the same frozen test repairs under ADD-first (used in our paper), reversed REMOVE-first, and two random-but-fixed orders preserves complete phenotype-group verifier responses across all distinct successful repairs. QuotientPO consistently outperforms Vanilla GRPO under all four mappings, confirming that its coverage advantage persists across reduction conventions [see also our response to Reviewer xfrh, Q4-Q5 on training-map robustness].
Deletion orderVanillaQuotientPOPaired Δ\Delta
Original0.5016±0.01290.5016 \pm 0.01290.5701±0.01980.5701 \pm 0.0198+0.0685±0.0168+0.0685 \pm 0.0168
REMOVE-first0.5839±0.04560.5839 \pm 0.04560.6510±0.01200.6510 \pm 0.0120+0.0671±0.0476+0.0671 \pm 0.0476
Random-fixed 10.5577±0.02470.5577 \pm 0.02470.6135±0.00550.6135 \pm 0.0055+0.0558±0.0212+0.0558 \pm 0.0212
Random-fixed 20.5824±0.02830.5824 \pm 0.02830.6444±0.01830.6444 \pm 0.0183+0.0619±0.0103+0.0619 \pm 0.0103

Revisions. We have refined the methodological details and robustness discussion in Sections 4.3 and 5.4, with detailed theoretical clarifications on factorized policies and verifier-preserving reduction in Appendices A.1–A.2, and comprehensive robustness analyses in Appendix F.7.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer KTZQ,

Thank you again for your thoughtful review and constructive feedback. We have provided detailed clarifications and additional experiments addressing your concerns.

We would greatly appreciate any further feedback or remaining concerns you may have after reviewing our responses. We would be happy to provide additional clarification if needed.

Best regards,

Paper 20414 Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response by Authors

We sincerely thank all the reviewers for their time, careful evaluation, and constructive feedback. The reviews focus on three main questions: (1) methodological novelty relative to existing diversity methods (xfrh), (2) theoretical validity and robustness of core reduction and modeling dependencies between reaction edits (KTZQ, xfrh), and (3) scientific significance and practical applicability of the proposed framework (hPQr, xfrh). We address these concerns through new experiments and manuscript revisions (highlighted in blue):

  • Methodological novelty: Redefining the unit of exploration. Our work addresses a fundamental challenge in scientific model repair: functional verification does not imply structural identification. Raw edit identity overcounts redundant solutions, whereas identical verifier outcomes can conceal distinct structural hypotheses. Beyond transition-based action redundancy [1], learned semantic partitions [2], and set-level diversity optimization [3], QuotientPO introduces an intermediate, verifier-preserving quotient that makes irreducible repair hypotheses the unit of exploration. We establish representation-invariant exploration as a principled objective, characterize its invariance condition, and prove that quotient entropy eliminates action entropy's representation-multiplicity bias at the unrestricted optimum. We further derive tractable kernelized Rényi-2 exploration with informative finite-sample signals and asymptotic local policy-gradient consistency (Sections 2-4 and Appendices A–B).

  • Soundness and robustness: Gains persist across reduction maps and phenotype-label counts. We clarify why a fixed, verifier-preserving reduction suffices for Theorem 4.1 without requiring NP-hard minimum-cardinality core extraction and how factorized Bernoulli policies learn coordinated multi-edit repairs through joint FBA rewards. New experiments show that QuotientPO's gains hold under stronger pairwise reduction, alternative deletion orders, independent training/evaluation maps, and varying phenotype-label counts (Sections 4.3, 5.4 and Appendices A.1-A.2, F.7, F.9).

  • Scientific significance: Phenotype-supervised GEM repair and biochemical hypothesis discovery. Metagenomics enables microbial genome reconstruction without cultivation, yet many microorganisms remain difficult to culture, leaving their growth phenotypes largely unknown. Traditional phenotype-guided GEM repair relies on organism-specific growth observations. We bridge this gap by learning transferable reaction corrections from phenotype supervision across organisms, without requiring gold-standard repaired GEMs. This enables repair generation and candidate prioritization from genomic and network evidence alone at test time. Beyond operational core counts, stoichiometric and FBA counterfactual analyses show that QuotientPO achieves higher joint recovery of distinct biochemical alternatives across four representative families (Introduction and Appendices F.8, G.2).

Additional optimization baselines. We further evaluate online rejection-sampling fine-tuning (xfrh Q1) and adapt SetPO (xfrh Q6) under matched training budgets. These comparisons support the benefits of policy-gradient optimization over verifier-filtered self-training and quotient-space exploration over raw-set diversity (Sections 5.2, 5.4 and Appendices D.3-D.4). We also provide their implementations and usage instructions in the updated supplementary code.

Broader ML contribution: From AI4Sci to Sci4AI. By revealing how partial identification in graph-structured model repair exposes a fundamental limitation of conventional exploration, QuotientPO demonstrates how scientific challenges can drive new ML formulations and principles, connecting AI4Sci with Sci4AI.

[1] Baram, Tennenholtz, and Mannor. Action Redundancy in Reinforcement Learning. UAI, 2021.

[2] T. Li et al. Jointly Reinforcing Diversity and Quality in Language Model Generations. arXiv:2509.02534, 2025.

[3] C. Li et al. SetPO: Set-Level Policy Optimization for Diversity-Preserving LLM Reasoning. ICML, 2026.

AuthorsOpenReview ↗

A Final Note on Our Rebuttal

Dear Reviewers and Area Chair,

As the author–reviewer discussion period draws to a close, we would like to thank you once again for your careful evaluation of our work.

We have taken the reviewers' concerns seriously and made substantial efforts to address them through targeted experiments, theoretical clarifications, and corresponding manuscript revisions.

While we have not yet received any feedback on our rebuttal, we sincerely welcome any remaining questions or concerns, particularly those that may affect the final assessment. We would appreciate the opportunity to address them if needed.

As the paper moves into the reviewer–AC discussion, we respectfully hope that the additional evidence and revised manuscript will be carefully considered alongside the initial reviews, especially where our responses directly address the original concerns.

Thank you for your time and consideration.

Best regards,

Paper 20414 Authors

AuthorsOpenReview ↗

A Final Update and Thanks

Dear Reviewers and Area Chair,

We would like to provide a brief update to our final note. In the final hours of the discussion period, Reviewers KTZQ and hPQr kindly provided their feedback, confirming that our responses had addressed their concerns and questions, respectively.

We are especially grateful to both reviewers for taking the time to revisit our work at this late stage and for their constructive feedback. We also sincerely appreciate the Area Chair's time and effort in overseeing the review process.

Thank you all again for your time and consideration.

Best regards,

Paper 20414 Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.