- 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 (), for example the same greedy procedure under 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 . 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
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.
| Metric | Method | 1–4 labels | 5 labels | 6–16 labels | All |
|---|---|---|---|---|---|
| GEMs | 1032 | 450 | 730 | 2212 | |
| Success@32 (%) | Vanilla GRPO | 22.71 | 15.48 | 12.69 | 17.93 |
| QuotientPO | 25.07 | 17.93 | 14.43 | 20.10 | |
| Successful Cores@32 | Vanilla GRPO | 0.59 | 0.45 | 0.41 | 0.50 |
| QuotientPO | 0.66 | 0.52 | 0.47 | 0.57 |
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.
| Method | Success@1 (%) | Success@32 (%) | Successful Cores@32 |
|---|---|---|---|
| Soft-SFT | |||
| RFT-Success | |||
| RFT-Improved | |||
| Vanilla GRPO | |||
| QuotientPO |
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 ( identity, bilateral coverage, E-value ) 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 order | Vanilla | QuotientPO | Paired |
|---|---|---|---|
| ADD-first | |||
| REMOVE-first | |||
| Random-fixed 1 | |||
| Random-fixed 2 |
| Reduction order | Vanilla | QuotientPO | Paired |
|---|---|---|---|
| Shuffled prior | |||
| Biological prior |
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.
Response to Reviewer xfrh (Part 1)
We thank the reviewer for recognizing our conceptual insight and comprehensive experiments.
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.