acceptodds
Under review as a conference paper at ICLR 2027

Measuring and Amplifying Self-Verification on Hard Problems

Abstract

An open question in scalable oversight is whether language models can assess solutions to problems that they cannot reliably solve. We test this by separating an assessor's informedness from its leniency. Informedness refers to a binary assessor's ability to distinguish correct from incorrect solutions. Leniency refers to its tendency to accept. Across six language models and seven benchmarks in mathematics, coding, and knowledge, we evaluate assessment on problems the model itself rarely solves (own solve rate ). We find that informedness on difficult problems strongly depends on task family: it is significantly higher than chance on mathematics and coding, but near the chance baseline on knowledge tasks. Next, we investigate committees of assessors and find that they can increase informedness compared to a single assessor. However, the gain is not automatic: with majority voting, the committee is worse than a single assessor in 6 of 18 settings. We therefore apply a simple mathematical model to predict committee behavior and select consensus thresholds. Taken together, our results identify task family, assessor behavior, and aggregation rule as key determinants of whether oversight remains effective beyond a model’s own solving capability.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Scalable oversight, LLM judges, self-verification, verifier models, large language models, AI safety, test-time compute

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_u2knOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether language models can verify solutions to problems they rarely solve themselves. It characterizes verifier behavior using informedness (Youden’s J) and leniency (L), distinguishing discrimination between correct and incorrect answers from the tendency to accept. Across six models and seven benchmarks, the authors find positive informedness on hard mathematics and coding problems, but no statistically significant departure from zero on hard knowledge questions. The paper also studies committees formed by repeated judgments from the same model. It develops mean-only and Beta approximations to predict committee performance and select voting thresholds. Both approaches improve informedness over majority voting on average, while the Beta approximation more accurately predicts committee operating curves.

Critical strengths
  1. The paper investigates a practically important question: whether language models can assess solutions to problems they rarely solve themselves. Its evaluation across six models and seven benchmarks reveals task-dependent differences in verification ability, providing useful evidence about when self-verification may support scalable oversight.

  2. The use of informedness and leniency provides an interpretable characterization of assessor behavior. By distinguishing discrimination between correct and incorrect solutions from the tendency to accept, this framework helps avoid misleading conclusions based on accuracy alone.

Critical weaknesses
  1. Limited methodological novelty. Informedness (Youden’s J) is an established metric, and its combination with leniency is an easy reparameterization of TPR and FPR. The committee analysis also builds on standard binomial aggregation and moment matching. While these tools provide a useful perspective, the paper does not sufficiently establish what methodological contribution goes beyond combining existing concepts.

  2. Unclear motivation and organization. The introduction is too brief to clearly explain the motivation, the limitations of existing work, and how the proposed approach addresses them. Although the introduction identifies limitations of prior studies, it does not sufficiently explain how the individual components of the proposed framework address these limitations or connect to one another. Besides, Section 3 introduces several concepts, including utility, the generation–verification gap, and randomized verdict flipping, before making their roles in the subsequent analysis clear. This organization makes the central argument difficult to follow and obscures the distinction between supporting analysis and the main methodological contribution. Some abbreviations are also used without being spelled out at first mention, further reducing readability.

  3. Limited comparison with existing methods. The related work discusses several verifier aggregation approaches, but the experiments mainly compare against simple baselines such as a single assessor and majority voting. Comparisons with relevant existing methods under matched computational budgets would help establish the practical value of the proposed approach.

Questions for authors
  1. The first limitation concerns benchmark-level aggregation mixing problems of different difficulty. Would stratifying questions by the model’s solve rate be sufficient to address this limitation?

  2. Do the authors provide any goodness-of-fit analysis showing how well the Beta distributions capture the class-conditional distributions of acceptance propensities, beyond the reported prediction errors for committee-level informedness and leniency?

  3. Can the authors provide theoretical guarantees on how Beta approximation and estimation errors affect committee performance predictions and threshold selection?

  4. The fitting and evaluation stages use different verdicts on the same proposals. Do the prediction accuracy and threshold-selection gains persist when calibration and evaluation use disjoint questions?

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer u2kn

We thank the reviewer for their thoughtful comments. We appreciate their recognition of the practical importance of studying self-verification beyond a model's own solving capability, and of the value of separating informedness from leniency to characterize assessor behavior. We have revised the analysis and presentation in response to their concerns, which we address below.

Responses to weaknesses:

  1. Methodological novelty. We agree that Youden's JJ, the informedness-leniency parameterization, and classical committee theory build on established statistical concepts. Our contribution lies in using these tools to study self-verification conditioned on a model's own per-question solving ability, characterize the resulting task-family differences, and predict and select committee operating points.

  2. Motivation and organization. We have reorganized the paper to distinguish the main empirical contributions from supporting theory. In particular, we moved the utility interpretation and generation-verification gap analysis to Appendix E. The main Methods section now focuses on experimental procedures, committee prediction, and threshold selection.

  3. Comparison with existing methods. We study homogeneous committees formed from repeated binary verdicts of one model. Within deterministic, monotone aggregation rules that depend only on the acceptance count, the problem reduces to selecting a threshold kk. We compare majority voting, mean-only and Beta-selected thresholds, and the best observed threshold as an optimistic reference (Table 2). These committee methods use the same nine verdicts per proposal at evaluation time, although calibration requires additional measurements. We also compare against randomized verdict flipping at matched leniency. Methods using heterogeneous verifiers or richer assessment signals are outside this setting, and we do not claim superiority over them.

Responses to questions

  1. Would stratifying by solve rate address the first limitation? Yes, it addresses the problem of benchmark-level averages mixing questions of different difficulty. We estimate per-question solve rates from 100 proposals and examine verification across difficulty bins (Figure 1) and cutoffs (Table 1). We also use JJ to measure discrimination independently of class prevalence. Stratification does not, however, eliminate differences in proposal format or error characteristics across tasks.

  2. Beta goodness-of-fit analysis. We have added a diagnostic comparing predicted and observed acceptance-count distributions in held-out nine-member committees (Section 4.2 and Appendix D.1). The Beta approximation achieves KL divergences of 0.028 for correct and 0.019 for unsuccessful proposals, compared with 1.428 and 1.709 for mean-only. Beta performs better in all 40 valid model-benchmark comparisons for each class. We also simulate verdicts under the fitted Beta model to estimate the divergence expected from finite sampling. The remaining discrepancy, particularly for correct proposals, indicates that the approximation is not exact. This diagnostic evaluates observable acceptance-count distributions rather than latent acceptance propensities directly.

  3. Are there theoretical guarantees? We do not establish guarantees for the fitted Beta approximation. Matching the first two moments does not uniquely determine the higher moments required for committee prediction, so accuracy of the Beta approximation requires additional assumptions. We instead evaluate it empirically using held-out verdicts, distributional goodness-of-fit, and threshold-selection performance, including on disjoint questions. Establishing guarantees under specific distributional assumptions remains an open direction.

  4. Do the gains persist on disjoint questions? Yes. As we also outline in our response to reviewer xqMj, we now include five-fold question-disjoint cross-validation (Section 4.2 and Appendix B.5). The mean-only and Beta-selected committees retain gains of ΔJ\Delta J = 0.046 [0.036, 0.058] and 0.043 [0.030, 0.057], respectively, compared with 0.011 [-0.008, 0.028] for majority voting. On held-out questions, Beta also predicts the operating curves more accurately than mean-only, with RMSEs of 0.121 versus 0.192 in JJ and 0.077 versus 0.199 in LL. Thus, the gains from threshold selection largely persist, although curve prediction becomes less accurate than in the same-proposal evaluation. Generalization across different question distributions remains untested and is an interesting direction for future work.

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

The paper studies whether a model can judge solutions to problems it cannot solve itself, splitting a binary assessor into J (discrimination) and L (tendency to accept). Across six models and seven benchmarks it reports that informedness stays above chance on hard math and coding but is at chance on hard multiple-choice knowledge, then studies committees of repeated assessor samples and a Beta moment-matching model for picking the consensus threshold.

Critical strengths
  1. Per-question accuracy estimation is the right design call, executed cleanly. Estimating proposer accuracy from 100 samples per question, rather than borrowing a benchmark aggregate, removes the confound that undermines most generation-vs-verification work. This is consequential because it isolates the actual signal of interest and makes the comparison between generation and verification far more interpretable than in prior work.

  2. The math is correct and the analysis is unusually careful. The unbiased second-moment estimator is a careful touch similar papers usually miss. Tables are internally consistent and the cutoff sensitivity analysis is useful. This is consequential because it gives strong confidence that the quantitative claims rest on solid ground, not on hand-waving or convenient approximations.

Critical weaknesses
  1. The headline task-family gap is largely confounded by how the question sets were constructed. The hard knowledge set sits at or below the random-guessing floor (mean solve rate 0.104 vs a 0.10 floor), so wrong answers look like right one: knowledge actually passes more wrong (61%) than right (56%)—while the math set sits ~100× above its floor, making failures visibly broken. With 84% of the family gap living in the rejection rate and no detectable difference in acceptance, the gap may reflect question-set construction rather than a genuine task-family effect. This is consequential because the paper’s central claim rests on this gap, yet the five-way outcome labels needed to diagnose it already exist in the pipeline and are never used.

  2. The assessor is never given a real chance to verify. All six models run with reasoning disabled or low (none high), two see proposals with the reasoning trace stripped, and there is no ablation, no cross-model cell, and no reference point of any kind (no stronger model, trained verifier, or human). So J = 0.21 on math has no scale, and “knowledge is at chance” cannot be separated from “knowledge is at chance when the judge is not allowed to think.” This is consequential because it leaves the paper unable to say whether the verifier’s poor performance on knowledge reflects a fundamental limitation or simply an underpowered assessor.

Questions for authors
  1. Can the study helps to imporve LLM performance? I didn't see this in the paper.
1 comment
AuthorsOpenReview ↗

Author Response to Reviewer XemL

We thank the reviewer for their thoughtful and detailed comments. We appreciate their recognition of the per-question accuracy estimation as an important experimental design choice, as well as the care taken in our mathematical derivations and quantitative analyses. We have incorporated their feedback into the revision and we respond to their concerns and questions below.

  1. Task-family gap and question construction. We agree that answer format and error characteristics may contribute to the observed gap. Under our hard-problem cutoff, mean solve rates are similar for mathematics (0.096) and knowledge (0.104), although the multiple-choice knowledge benchmarks have explicit random-guessing baselines, unlike open-ended mathematics. We note however that the two knowledge benchmarks also have different guessing baselines (0.25 for GPQA-Diamond and 0.10 for MMLU-Pro), yet informedness is near zero on each separately (J = -0.028 and -0.081).

In the updated manuscript we include additional analysis that address several potential confounds. In particular, following the reviewer's suggestion we report the breakdown of unsuccessful proposals into incorrect, truncated, and no-answer outcomes. Completed-incorrect proposals account for at least 99% of unsuccessful proposals in each task family. Restricting the negative class to these proposals leaves the mathematics--knowledge gap essentially unchanged (ΔJ\Delta J = 0.267 versus 0.269). We also add an assessor-facing response-format control (ΔJ\Delta J = 0.048 [-0.035, 0.131]) and an frequent-error-acceptance analysis showing that assessors preferentially accept frequently generated incorrect knowledge answers (FPR difference 0.125 [0.037, 0.214]). These results indicate that incomplete proposals do not explain the gap, while preferential acceptance of frequently generated incorrect knowledge answers may contribute to it. However, the effects of task content and proposal elicitation remain unresolved.

  1. Assessor reasoning effort. We match proposer and assessor inference configurations to study verification relative to each model's own solving ability under the same settings. We agree that this does not characterize the models' maximum verification capability. Higher assessor reasoning could increase verification capability but should be tested against proposers with the same reasoning effort. This is now recognized as future work in the conclusion. Assessors receive the full visible response content, including explanations; only separately returned reasoning traces or summaries are excluded (Appendix B.1). We have clarified the inference-configuration limitation in Section 4.1 and acknowledge that stronger-verifier comparisons remain outside our experiments.

  2. Implications for improving LM performance (Question 1). Our results identify conditions under which verification can improve LM-based systems. As derived in Appendix E, for a fixed question with positive informedness, rejection sampling increases the probability that an accepted proposal is correct, provided the original success probability is between zero and one and acceptance is possible. Our committee results further show how aggregation can improve informedness and how acceptance thresholds can be selected. For LM self-improvement, one possible application is filtering self-generated solutions before using them as training data, thereby reducing the proportion of incorrect examples.

Reviewer_xqMjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem that if LLMs can actually recognize correct solutions to problems that they are barely able to solve themselves independently. Experiments with various setups were conducted to test different ways of "judgements" and the results show that they manifest "better-than-change" judgement on selected math and coding problems, but no clear advantage over chance on knowledge tasks. Essembling judgements can often improve the result but the sampling strategy is also less trivial (e.g., than simply following the majority).

Critical strengths
  • The diff between math/coding and knowledge appears across all the models tested, which provides some useful evidence against the assumption that verification is always easier than solving the problem.
  • The voting experiments offer a practical lesson as well, e.g, more judgements do not necessarily improve verification, and more thoughtful design is needed to get it right.
Critical weaknesses
  • The analysis across domains could be conducted indepth, e.g, it's still unclear how much of the math-vs-knoweldge gap reflects genuine verification ability rather than the differences in e.g., the format or the completeness of the answer itself, considering the different nature of different tasks.
  • The voting mechenism is fitted and tested on the same solutions, but does not imply that the chosen strategy generalizes on new problems, which is perhaps important to its practical value.
Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer xqMj

We thank the reviewer for their thoughtful comments. We appreciate their recognition of the value of our work and the meaningful feedback. Much of the feedback has been incorporated into the revised manuscript, and we address the reviewer's two main concerns below.

  1. Task-family differences and confounding. We agree that response format and proposal characteristics may contribute to the mathematics--knowledge gap. We have added three analyses in Section 4.1. First, changing the assessor-facing multiple-choice presentation substantially changes TPR and TNR, but produces no clear change in informedness (ΔJ\Delta J = 0.048 [-0.035, 0.131]). Second, restricting the negative class to completed-incorrect proposals leaves the gap essentially unchanged (ΔJ\Delta J = 0.267, compared with 0.269 originally). Third, assessors accept the most frequently generated incorrect knowledge answers more often than other incorrect answers (FPR difference 0.125 [0.037, 0.214]). These findings rule out incomplete responses as the primary explanation and provide further evidence about error characteristics, although they do not isolate the effect of task family from proposal format and elicitation.

  2. Generalization to unseen questions. We thank the reviewer for this helpful suggestion. We have now added five-fold question-disjoint cross-validation (Section 4.2), selecting committee thresholds on calibration questions and evaluating them on held-out questions. The mean-only and Beta-selected committees retain informedness gains of ΔJ\Delta J = 0.046 [0.036, 0.058] and 0.043 [0.030, 0.057], respectively, compared with 0.011 [-0.008, 0.028] for majority voting. Thus, the threshold-selection gains largely persist on unseen questions, but we note that generalization to different question distributions remains untested.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.