acceptodds
Under review as a conference paper at ICLR 2027

CDRL: Certification-Driven Reinforcement Learning for Neutrino Flavor Model Discovery

Abstract

Many scientific discovery problems require searching combinatorial hypothesis spaces under complex domain constraints. Reinforcement learning (RL) offers a promising approach, but existing methods rely on scalar rewards that provide limited information about why candidate solutions fail, leading agents to repeatedly explore invalid regions. We introduce Certification-Driven Reinforcement Learning (CDRL), a framework that leverages structured feedback from symbolic reasoning tools. When a candidate violates domain constraints, these tools produce certificates identifying the actions responsible for failure. CDRL converts these certificates into reusable constraints that eliminate classes of invalid solutions and guide exploration toward valid regions. We evaluate CDRL on neutrino flavor model discovery in theoretical particle physics, where the hypothesis space exceeds possible models, and compare it with the state-of-the-art RL approach previously used for this task. Across three theory spaces, CDRL achieves up to 1.95 higher valid model rates and up to 6.33 higher neutrino model rates while evaluating up to 4 fewer candidates. We further extract 40 interpretable rules from search trajectories using a post-hoc decision-tree framework and show that reusing them as soft constraints yields gains of up to 2 in valid model rates and 3 in neutrino model discovery across all three theory spaces. These results suggest that CDRL uncovers reusable structure in combinatorial search spaces and provides a general framework for scientific model discovery.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neuro-Symbolic AI, Scientific Discovery, Combinatorial Search, Symbolic Reasoning, Constraint Learning, Monte Carlo Tree Search, SAT Solving, Particle Physics, Neutrino Flavor Model Discovery

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VLaoOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CDRL, which combines neural-guided MCTS with symbolic reasoning and reusable failure certificates for neutrino flavor-model search. Failed candidates produce constraints that are propagated to future search through SAT/BCP. The method is evaluated on three theory spaces and compared with AMBer and several controlled ablations.

Critical strengths

The mechanism is well motivated. Instead of reducing an expensive scientific evaluation to a scalar reward, the method reuses structured failure information to prune future search.

The component ablations are convincing. Removing certificates, MCTS, or the neural component produces substantially different degradation patterns, and these experiments are reported across multiple random seeds.

Critical weaknesses

The main limitation for me is novelty positioning. MCTS, neural guidance, SAT propagation, and conflict-style reusable constraints are individually established ideas, so I see the contribution primarily in their integration for this scientific-search setting rather than as a fundamentally new RL paradigm.

I would also interpret the AMBer comparison more cautiously. CDRL uses substantially fewer terminal physics evaluations, which is a meaningful result, but the two systems do not use matched execution configurations, and the paper explicitly does not treat its wall-clock comparison as matched compute.

Questions for authors

Can the authors provide a smaller matched-compute comparison with AMBer or an equivalent scalar-reward baseline using the same hardware, caching, parallelism, and physics-evaluation stack? This would make the efficiency claim substantially stronger.

3 comments

Q1. Novelty

A: We uploaded a revised PDF with all changes marked in green. Section, figure and table numbers below refer to the revised version. We agree that MCTS, neural guidance, SAT propagation and conflict-style constraints are each established. What CDRL introduces is the mechanism that connects them. To the best of our knowledge, hard constraints learned during search have not been used inside the DNN+MCTS setting of AlphaZero, which is widely used for discovery because it learns and plans at the same time [1, 2, 3]. Earlier work used MCTS to improve SAT solvers [4, 5], but CDRL works in the other direction and uses SAT reasoning to prune a learned MCTS search.

For example, Neo[6] looks for one program that meets a given specification and finds conflicts by reasoning over logical specifications of the program’s components. It uses them to prune a backtracking search guided by a fixed, pretrained model, with no planning and no learning during search. CDRL looks for many inequivalent models under an expensive evaluator. A rank failure only becomes visible after the physics pipeline has evaluated a complete model, so CDRL learns its clauses from these evaluations. It removes both invalid and redundant models, and its clauses act inside a self-play loop where MCTS plans with a policy-value network that keeps training. Deduction-guided RL [7] uses deductions to update the policy, which only lowers the probability of a bad choice, while CDRL clause removes the choice. Oracle-guided synthesis [8] passes constraints from an oracle to a solver that looks for one solution, without a learned policy or planning. Symbolic advice for MCTS [9] uses a formula fixed by the user before search, which only biases selection within a bounded horizon and never changes. In CDRL, the clause set grows during search from what the evaluator reveals, every clause is shared across workers, and a canonical-representative rule keeps them sound under parallel sharing. The setting is also new in physics. RL has been used for model building [10, 11, 12], and SAT solvers have searched string vacua with fixed constraints [13], but we are not aware of work that learns constraints from a physics evaluator during search. Bringing conflict-driven learning to flavor model building is itself new, and the mapping is non-trivial. AMBer [10] uses the same physics pipeline but learns from failures only through a scalar reward. CDRL adds the missing layer which is a certificate analyzer that extracts the choices responsible for each failure or redundancy, and a SAT encoding that turns them into clauses the search enforces.

This mechanism, not the individual components, accounts for the gains. Under a matched budget and hardware, CDRL’s valid model rate is 2.2x, 2.8x and 2.1x that of AMBer, and its neutrino model rate is 10x, 19x and 4.5x. Most of this gain comes from the certificates. Removing them while keeping the same planner, network and reward drops the valid rate from 28.1% to 14.9%, 26.9% to 16.2%, and 18.2% to 5.4%, close to AMBer’s level. Adding the same certificates to PPO helps only modestly, so the gains come from combining learned constraints with learning and planning, which is what CDRL introduces.

References

[1] Silver et al., Mastering the Game of Go without Human Knowledge, Nature 2017.

[2] Fawzi et al., Discovering Faster Matrix Multiplication Algorithms with Reinforcement Learning, Nature 2022.

[3] Mankowitz et al., Faster Sorting Algorithms Discovered using Deep Reinforcement Learning, Nature 2023.

[4] Schloeter, A Monte Carlo Tree Search Based Conflict-Driven Clause Learning SAT Solver, INFORMATIK 2017.

[5] Keszocze et al., Improving SAT Solving Using Monte Carlo Tree Search-based Clause Learning, Advanced Boolean Techniques 2020.

[6] Feng et al., Program Synthesis using Conflict-Driven Learning, PLDI 2018.

[7] Chen et al., Program Synthesis Using Deduction-Guided Reinforcement Learning, CAV 2020.

[8] Polgreen et al., Satisfiability and Synthesis Modulo Oracles, VMCAI 2022.

[9] Busatto-Gaston et al., Monte Carlo Tree Search Guided by Symbolic Advice for MDPs, CONCUR 2020.

[10] Baretz et al., Towards AI-assisted Neutrino Flavor Theory Design, Communications Physics 2026.

[11] Halverson et al., Branes with Brains: Exploring String Vacua with Deep Reinforcement Learning, JHEP 2019.

[12] Harvey and Lukas, Quark Mass Models and Reinforcement Learning, JHEP 2021.

[13] Faraggi et al., Satisfiability Modulo Theories and Chiral Heterotic String Vacua with Positive Cosmological Constant, Phys. Lett. B 2021.

Q2. Can the authors provide a matched-compute comparison with AMBer using the same hardware, caching, parallelism, etc.?

A: We thank the reviewer for this suggestion and have run both comparisons. First, we reran AMBer's PPO agent with four seeds on the same hardware, with the same evaluation cache and number of parallel workers, with the same evaluation budget per run (1M, 1M and 6M). Second, our No symbolic ablation is an equivalent scalar-reward baseline that matches CDRL on every part of the setup, i.e., hardware, caching, parallelism, physics stack, planner, network, reward and budget. It differs only in that it learns from the scalar reward alone, without certificates.

Search spaceMethodEvaluations per runValid models (%)Neutrino models (%)Min. parameters
A4 x Z4Full CDRL1M28.1 +/- 3.30.10 +/- 0.015
No symbolic (scalar reward only)1M14.9 +/- 3.10.01 +/- 0.027
AMBer PPO1M12.5 +/- 5.00.01 +/- 0.015
PPO + SAT & certificates1M13.9 +/- 4.10.02 +/- 0.005
A4 x ZNFull CDRL1M26.9 +/- 3.70.19 +/- 0.055
No symbolic (scalar reward only)1M16.2 +/- 2.10.03 +/- 0.017
AMBer PPO1M9.7 +/- 3.50.01 +/- 0.015
PPO + SAT & certificates1M11.0 +/- 3.10.01 +/- 0.015
T19 x Z4Full CDRL6M18.2 +/- 2.10.09 +/- 0.024
No symbolic (scalar reward only)6M5.4 +/- 4.90.04 +/- 0.025
AMBer PPO6M8.8 +/- 3.60.02 +/- 0.015
PPO + SAT & certificates6M11.3 +/- 2.60.03 +/- 0.014

Under these matched conditions, CDRL's valid model rate is 2.2x, 2.8x and 2.1x that of AMBer, and its neutrino model rate is 10x, 19x and 4.5x. The fully matched scalar-reward baseline shows where this gain comes from. With the same stack and budget but no certificates, the valid rate falls to 14.9%, 16.2% and 5.4%, close to AMBer's level. Certificates therefore account for most of the improvement, not differences in execution configuration. Adding the SAT layer and certificates to PPO also helps (12.5% to 13.9%, 9.7% to 11.0%, 8.8% to 11.3%), but the gain is much larger with MCTS, where each clause prunes the lookahead at every node it reaches. We have added these matched comparisons to Section 5.2, Figure 2(a,b) and Table 6 (Appendix G.4).

Summary of additional results

Beyond the points above, the revision adds several new experiments and analyses.

  • Certificate-type ablations and feasibility-check baselines. Learning rank clauses clearly beats checking rank on every model without learning from failures (e.g., 17.9% to 26.9% valid on A4 x ZN), learned symmetry clauses add to this, and the full system is best in every space. Even without the network and MCTS, adding learned clauses to a SAT search finds the first neutrino models and lowers the minimum parameter count from 12 to 7. See Section 5.3, Figure 2(c,d) and Table 7.
  • Table 1 gives four-run means and Table 2 the best run, following AMBer's convention, and both captions now state this explicitly.
  • We state the invariant that keeps one representative of every equivalence class available for any number of workers and any synchronization order, and verify it exhaustively for the moduli in our theory spaces (Appendix B.3).
  • We discuss why the analyzer carries over unchanged across the three theory spaces and which properties a new domain needs, such as molecule design and crystal structure search (Section 6).
  • A discussion of why many distinct viable models matter for upcoming tests of flavor models (DUNE, T2HK, neutrinoless double beta decay). See Appendix D.

We thank the reviewer for the positive assessment and the helpful suggestions. We hope the matched-compute comparison, the clarified novelty positioning and the additional experiments above resolve the remaining concerns, and we would be grateful if the reviewer would consider raising their score to further support the paper.

Dear reviewer, thank you again for your positive review. As the discussion period is ending soon, we wanted to check whether our response resolved your questions. We clarified the novelty positioning and added the requested matched-compute comparison: an AMBer rerun and our No symbolic scalar-reward baseline, both on the same hardware, cache, parallelism and evaluation budget as CDRL. If any concerns remain, we would be glad to address them. Otherwise, we would be grateful if you would consider raising your score.

Reviewer_p23KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper combines neural search with physics checks to find neutrino models. When a check identifies a structural failure, the system saves a rule to avoid similar failures later. It also uses symmetry rules to avoid searching equivalent models.

Critical strengths
  1. The paper addresses a useful problem. Physics evaluations are expensive, so avoiding repeated failures can make the search more efficient.

  2. The comparison with certificate learning removed supports the approach. In one setting, adding certificates increases the success rate from 0.03% to 0.19%. Improvements are also reported in the other settings.

  3. The authors use rules extracted from search as suggestions, rather than treating them as proven physics constraints. This is a reasonable choice.

Critical weaknesses
  1. The main efficiency claim is hard to verify. For the largest setting, the average success rate and evaluation budget suggest about 5,400 successful candidates, but the paper reports 7,019 distinct models. These numbers may come from different runs, but this is not explained. Consistent counts and budgets are needed to support the claim of finding more models with four times fewer evaluations.

  2. The shared pruning rules need a clearer explanation. When two charge assignments describe the same physics, the search must keep one of them. The paper does not explain how different workers agree on which one to keep when sharing rules. This matters because conflicting choices could remove possibilities that should remain available.

  3. The results support a narrower claim than the paper suggests. All three settings belong to the same physics task, and the method depends on checks written for that task. The results are useful, but they do not yet show how well the approach would work in other scientific problems.

Questions for authors
  1. Can the authors report success rates, distinct-model counts, and total evaluations for the same runs?

  2. How do workers ensure that their symmetry rules always leave an equivalent model available?

3 comments

Q1. Table 1 vs Table 2 numbers.

A: We uploaded a revised PDF with all changes marked in green. Section, figure and table numbers below refer to the revised version. We thank the reviewer for checking this carefully. Table 1 and Table 2 use different aggregation. Table 1 reports the mean and standard deviation over four runs of each per-evaluation rate. Table 2 follows AMBer’s reporting convention and gives the number of unique inequivalent neutrino models in the best run of each method, after AMBer’s filtering. The best run has a higher rate than the four-run mean, which explains the gap the reviewer notes. For the same best runs, CDRL used 1M, 1M and 6M evaluations and found 1,090, 2,343 and 7,019 distinct neutrino models. This corresponds to 0.109%, 0.234% and 0.117% distinct models per evaluation, compared with four-run means of 0.10%, 0.19% and 0.09%. AMBer’s best runs used 4M, 4M and 24M evaluations. We now state this explicitly in both captions. We also reran AMBer with four seeds on the same hardware and the same budget per run as CDRL. Under these matched conditions, CDRL’s mean valid model rate is 2.2x, 2.8x and 2.1x AMBer’s, and its mean neutrino model rate is 10x, 19x and 4.5x, so the efficiency gain does not depend on how the counts in Table 2 are aggregated.

Q2. How do workers ensure that shared symmetry rules always leave an equivalent model available?

A: We thank the reviewer for raising this. Workers never have to agree on which model to keep, because the choice is fixed by the equivalence class itself rather than by the order of evaluation or by which worker sees a model. When a worker evaluates a model, the certificate analyzer computes the full equivalence class of its charge assignment, i.e., every reduction and embedding into the allowed ZN groups, together with their negations. Within this class, the analyzer selects one canonical member by a fixed rule, the smallest (N, charge tuple) in lexicographic order, and emits clauses only for the other members. Since every worker that meets any member of the class computes the same class and the same canonical member, all workers block the same set of models and never block the canonical one. For example, two workers that evaluate Z4[q] and Z4[-q] before synchronizing both keep the same member and emit the same clause. Sharing rules only takes the union of these sets, so for any number of workers and any order of synchronization, the canonical member of every class remains available. Rank certificates need no such rule, since each one excludes only models that are invalid for every charge assignment.

We verified this invariant exhaustively for the moduli used in our theory spaces. For every charge assignment, the set of models blocked is the full class minus the same canonical member, whichever member is evaluated first. We now state the invariant explicitly in Appendix B.3.

Q3. Do the results support claims beyond this one physics task?

A: We agree that all our experiments are in one scientific class. Within this class, however, the checks are not tied to one setting. The three theory spaces use very different non-Abelian groups (A4 and T19), different numbers of flavons, and either a fixed Z4 or a choice of ZN from 2 to 10, so they draw on different mixes of certificates. Traditionally, you may have a different physicist working on a T19 problem compared to an A4ZN problem. For us, the same analyzer runs on all three without change, because its rules depend only on general facts. An added symmetry can only remove mass terms, which gives the rank certificates, and the selection rule of a ZN symmetry is preserved under reduction, embedding and negation, which gives the symmetry certificates. Neither fact is specific to A4 or to these particles, so the same analyzer applies to other discrete flavor groups and to other sectors such as quark masses.

Beyond physics, the part of CDRL that depends on the domain is deliberately small. Algorithm 1 needs a verifier and an analyzer whose clauses satisfy the soundness condition in Appendix B.3. The planner, the SAT layer, clause sharing and rule extraction carry over unchanged. This matches how conflict-driven methods work in general. Natural next applications have the same structure, ie., a discrete design space, an expensive evaluator, and known relaxations or equivalences that give sound clauses. Examples include molecule design, where valence rules and graph isomorphism play the roles of our rank and symmetry rules, and crystal structure search, where space-group symmetry plays the same role. We have added this discussion to Section 6 and Appendix B.5.

Summary of additional results

Beyond the points above, the revision adds several new experiments and analyses.

  • We reran AMBer's PPO agent with four seeds on the same hardware, cache, parallelism and evaluation budget as CDRL. CDRL's valid model rate is 2.2x, 2.8x and 2.1x that of AMBer, and its neutrino model rate is about an order of magnitude higher, so the gain is larger than in the original comparison. See Section 5.2, Figure 2(a,b) and Table 6.
  • We also ran PPO with our SAT layer and certificates. Certificates help PPO only modestly (e.g., 9.7% to 11.0% valid on A4 x ZN), so the gain comes from combining learned constraints with MCTS and the network. See Section 5.2 and Table 6.
  • We also performed certificate-type ablations and feasibility-check baselines. Learning rank clauses clearly beats checking rank on every model without learning from failures (e.g., 17.9% to 26.9% valid on A4 x ZN), learned symmetry clauses add to this, and the full system is best in every space. Even without the network and MCTS, adding learned clauses to a SAT search finds the first neutrino models and lowers the minimum parameter count from 12 to 7. See Section 5.3, Figure 2(c,d) and Table 7.
  • Positioning against conflict-driven synthesis (Neo), deduction-guided RL, synthesis modulo oracles and symbolic-advice MCTS, now in Section 2 and Appendix A.2.
  • A discussion of why many distinct viable models matter for upcoming tests of flavor models (DUNE, T2HK, neutrinoless double beta decay). See Appendix D.

We thank the reviewer for the positive assessment and the constructive feedback. We hope the clarifications and additional experiments above resolve the remaining concerns, and we would be grateful if the reviewer would consider raising their score to further support the paper.

Dear reviewer, thank you again for your positive review. As the discussion period is ending soon, we wanted to check whether our response resolved your questions on the Table 1 and 2 reporting and on how parallel clause sharing keeps an equivalent model available. If any concerns remain, we would be glad to address them. Otherwise, we would be grateful if you would consider raising your score.

Reviewer_GQf4OpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper proposes Certification-Driven Reinforcement Learning (CDRL), a neuro-symbolic framework for efficiently exploring large combinatorial scientific hypothesis spaces.
  • The central motivation is that conventional reinforcement learning typically relies on scalar rewards to evaluate candidate solutions. When a candidate violates scientific constraints, a negative reward indicates failure but does not explicitly identify which decisions caused it. Consequently, the agent may repeatedly explore similar invalid solutions.
  • CDRL addresses this limitation by incorporating symbolic certificates that identify the assignments responsible for structural failures or redundant solutions. These certificates are converted into reusable logical constraints, allowing the agent to eliminate entire classes of invalid or equivalent candidates during subsequent exploration.
  • The authors demonstrate the framework on neutrino flavor model discovery, a theoretical particle physics problem involving more than 10^26 candidate models.
Critical strengths

S1. The paper addresses a challenging and practically relevant scientific search problem. Neutrino flavor model discovery involves an enormous combinatorial search space and computationally expensive physics evaluations.

Critical weaknesses

W1. The methodological novelty of CDRL appears relatively limited.

Its central idea of converting structured failure feedback into reusable constraints has substantial precedents in conflict-driven synthesis and neuro-symbolic reinforcement learning. In particular, Feng et al. (PLDI 2018) [1] introduced Neo, which identifies the root causes of unsuccessful candidates and learns logical constraints that eliminate entire classes of infeasible solutions. Chen et al. (CAV 2020) [2] further combined reinforcement learning with deductive reasoning, leveraging feedback about infeasible states to improve exploration. Related principles have also been studied in oracle-guided synthesis (Polgreen et al., VMCAI 2022) [3], while Busatto-Gaston et al. (CONCUR 2020) [4] investigated incorporating SAT/QBF-based symbolic guidance into MCTS. Although CDRL differs from these approaches by dynamically incorporating domain-specific certificates into neural-guided MCTS for scientific model discovery, the paper does not sufficiently establish what fundamentally new algorithmic mechanism this integration introduces.

Moreover, the proposed certificates are derived from explicitly designed checks for mass-matrix rank deficiency and known symmetry equivalences, rather than a general-purpose procedure for automatically extracting constraints from arbitrary scientific evaluators. Given that AMBer (Baretz et al., 2026) [5] already established the underlying neutrino-model discovery framework, the primary contribution appears to lie in integrating established search and reasoning techniques into this application.

The authors should more carefully differentiate CDRL from the aforementioned approaches and clarify the necessity of dynamically learned certificates relative to simpler alternatives, such as enforcing known symmetry constraints or structural feasibility checks directly during search.

  • [1] Feng, Y., Martins, R., Bastani, O., and Dillig, I. (2018). “Program Synthesis using Conflict-Driven Learning.” Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI)
  • [2] Chen, Y., Wang, C., Bastani, O., Dillig, I., and Feng, Y. (2020). “Program Synthesis Using Deduction-Guided Reinforcement Learning.” Proceedings of the 32nd International Conference on Computer Aided Verification (CAV)
  • [3] Polgreen, E., Reynolds, A., and Seshia, S. A. (2022). “Satisfiability and Synthesis Modulo Oracles.” Proceedings of the 23rd International Conference on Verification, Model Checking, and Abstract Interpretation (VMCAI)
  • [4] Busatto-Gaston, D., Chakraborty, D., and Raskin, J.-F. (2020). “Monte Carlo Tree Search Guided by Symbolic Advice for MDPs.” Proceedings of the 31st International Conference on Concurrency Theory (CONCUR)
  • [5] Baretz, J. B., Fieg, M., Ganesh, V., Ghosh, A., Knapp-Pérez, V., Rudolph, J., and Whiteson, D. (2026). “Towards AI-assisted Neutrino Flavor Theory Design.” Communications Physics, 9, Article 227

W2. Insufficient evidence of the scientific significance of the additional discovered models

While neutrino flavor model discovery is an important scientific problem, the paper does not sufficiently establish the scientific significance of discovering a larger number of candidate models. The main experimental improvements are measured by the number and proportion of models satisfying predefined criteria χ2≤10\chi^2 \leq 10 and np≤7n_p \leq 7, inherited from AMBer (Baretz et al., 2026). Although CDRL substantially increases the number of qualifying models, it does not improve the minimum number of free parameters found in any of the three theory spaces (Table 1). More importantly, it remains unclear whether the additional models provide meaningful scientific insights, such as previously unexplored model families, distinct physical predictions, or stronger agreement with experimental observations beyond the existing evaluation criteria. While Appendix D reports broader coverage of symmetry groups and presents an example discovered model, these results do not fully establish the additional scientific value of the discoveries.

Moreover, the evaluation relies on simplifying physical assumptions, including predefined flavon vacuum expectation values, so satisfying the reported criteria does not necessarily establish full physical viability. I encourage the authors to provide a more detailed analysis of the newly discovered models, particularly their physical distinctiveness, predictive implications, or validity under additional theoretical and experimental constraints. As currently presented, the results convincingly demonstrate improved computational search efficiency, but the broader scientific significance of the additional discoveries remains insufficiently substantiated.

W3. Insufficient comparisons with simpler constraint-aware search methods

The experimental comparisons do not sufficiently establish the necessity of the proposed certification-driven learning mechanism. The three certificate types primarily encode known structural or symmetry properties, including mass-matrix rank deficiency, Abelian charge reduction/embedding, and charge negation. Such properties could potentially be incorporated directly into the search process through predefined constraints, symmetry canonicalization, or inexpensive structural feasibility checks. However, the paper does not compare CDRL against these simpler constraint-aware alternatives. Moreover, the No symbolic ablation removes all certificate types simultaneously, making it difficult to determine which mechanisms contribute most to the observed improvements. I encourage the authors to include comparisons with static constraint enforcement or symmetry-breaking baselines, along with certificate-type-specific ablations.

W4. Limited comparability and statistical rigor of the main experimental evaluation

The comparison with AMBer involves several substantial differences beyond the proposed certificate mechanism, including the underlying learning algorithm (PPO versus neural-guided MCTS), model construction procedure, reward design, and parallel execution configuration. Furthermore, the main comparison uses previously published AMBer results rather than repeated evaluations under matched computational settings. Although the No symbolic ablation partially isolates the benefit of certification-driven pruning, the reported improvements over AMBer remain difficult to interpret under these differences.

In addition, CDRL is evaluated using one-quarter of AMBer's terminal-evaluation budget, and the baseline results do not include uncertainty estimates comparable to CDRL's four-run statistics. The relationship between the discovered-model counts in Table 2 and the mean discovery rates in Table 1 is also unclear. I encourage the authors to clarify their result aggregation and provide budget-matched discovery curves, comparable statistical analyses, and more controlled computational comparisons.

Questions for authors

See weaknesses above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.