- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
EviGRAD addresses errors in test-based rewards for reinforcement learning in code generation through selective verification and conservative reward correction. A learned risk model prioritizes rollout groups for judge-guided counterexample generation, and validated failing tests reduce the rewards of candidates that passed all original tests. A separate evidence-guided repair or resampling procedure handles low-reward groups with no within-group reward variation. The paper reports improvements over pass-rate GRPO on CodeContests and LiveCodeBench, using approximately one-third of the additional verification resources required by always-verify.
Critical strengths
None
Critical weaknesses
- I find the methodological advance over prior test-generation and verification approaches insufficiently established. The additional logistic-regression risk model offers a mechanism for prioritizing verification, but its practical value remains unclear, particularly because its features already depend on potentially expensive judge outputs. The experiments lack matched-budget downstream comparisons against strong test-augmentation methods and simpler selection strategies, such as random verification or seed-only selection. Consequently, the reported results do not establish whether EviGRAD’s additional machinery provides meaningful benefits beyond obtaining more informative tests and incorporating additional judge supervision.
Questions for authors
- Why is HardTests/HardTestGen neither discussed nor included as a baseline? It already investigates generating and validating challenging tests for use as RL verifiers, making it directly relevant to the claimed contribution. A compute-matched comparison using the same base model and training problems, with test-generation and verification costs included, is needed to establish what EviGRAD adds beyond improving test coverage. HardTestGen, ICLR 2026
- What gains are specifically attributable to the learned risk model and one-sided reward correction? The current comparisons do not adequately separate these contributions from access to additional tests, judge supervision, and repair/resampling. In particular, matched-budget downstream comparisons against random verification, seed-only selection, and standard reward recomputation using augmented tests are missing. The full method’s improvement over the degenerate-low-only branch also remains statistically unresolved in Table 15 (p = 0.19). Consequently, the experiments do not establish whether the additional design is necessary to obtain the reported gains.
- Could the authors reconcile Tables 4 and 14? Equation 4 defines an over-reward event as and . With the stated population of 29,900 candidates and , Table 4 reports 1,040 positives for , whereas Table 14 reports 3,140 over-reward cases in the corresponding positive-advantage stratum. For , the corresponding counts are 400 versus 1,430. Under the stated definitions, these counts should agree. Do the tables use different candidate populations or label definitions, or are these reporting errors? This discrepancy directly affects the calibration analysis underlying the method’s motivation.
- Why only one old model? This just does not make sense to me.