acceptodds
Withdrawn from ICLR 2027

Knowing Is Not Choosing: What Explicit Verification Adds Beyond Generative Preference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating a correct answer does not ensure that a language model will select it. We track factual competence through three stages: whether a correct candidate is generated, how candidates are ranked, and which answer is ultimately returned. Pre-generation readouts predict factual recall beyond exposure and identity features across three model families and, without refitting, which questions sampling will cover; however, they add little about whether an available correct answer is ultimately returned. Explicit verification with improves within-question ranking over mean log-likelihood in Gemma, Qwen3, and Llama, raising pair-weighted AUROC by –. In a prospectively specified Gemma cohort, verification improves plurality accuracy by points under a frozen stricter criterion and by points under blinded semantic labels. Against chat-template likelihood, a stronger generative baseline, the gain remains points ( CI ), concentrated in relations with strong common-answer priors and dependent on the entity. In two 27B Qwen models, verification exceeds chat-template likelihood by – AUROC; masking the entity removes this ranking advantage, while plurality gains remain at most points. The measured gain also depends on the correctness criterion: recall-oriented reference matching credits option lists favored by likelihood and reduces the plurality gain over mean log-likelihood to points, whereas human labels give .

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language models, factual knowledge, self-verification, inference-time sampling, candidate selection, confidence estimation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5eUhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the discrepancy between an LLM's ability to generate a correct answer and its ability to select the correct answer. The paper decomposes the factual question-answering process into three stages: whether the model generates a correct answer, how it ranks candidate answers, and how it selects the final answer. The paper especially compares the effects of selecting the final answer based on generation likelihood versus explicit verification. Through empirical evaluations across multiple frontier model families, the paper demonstrates that explicit verification methods improve ranking and plurality accuracy over standard mean log-likelihood and chat-template likelihood baselines. The paper also explores how evaluation metrics and entity masking affect the performance gains.

Critical strengths
  1. The paper addresses an interesting diagnostic research question: why LLMs have difficulty directly selecting the correct answer even if their generated candidates cover it. By examining candidate coverage, candidate ranking, and final answer selection, the paper provides a useful perspective for revisiting this problem.
  2. The paper provides extensive and comprehensive experiments to study the raised problem. The analysis covers multiple model families across Gemma, Llama, and Qwen, which makes the study generalizable. Additional analyses also help in understanding when verification improves ranking and whether those improvements translate into better final answers.
Critical weaknesses
  1. The presentation of this paper is disorganized and hard to follow. After a brief motivation, the introduction quickly shifts to detailed empirical results without fully explaining the background, research gap, or experimental setup. The discussion also moves between several baselines and correctness criteria, making it difficult to follow which comparisons support which conclusions. Overall, the paper's writing and expression do not meet this venue's standards.
  2. The paper is mostly analytical and observational, but does not provide
  3. The paper lacks an identifiable novel contribution. Gekhman et al. [1] already compare likelihood with P(True)P(\mathrm{True}) and discuss why better pairwise ranking need not improve final selection. Although the paper discusses this in related work and adds more controlled experiments and observations, it does not clearly reveal a novel scientific insight from these additions.

References: [1] Gekhman, Zorik, et al. "Inside-out: Hidden factual knowledge in llms." COLM 2025.

Questions for authors
  1. Can the authors clarify the conceptual structure of the paper and the connection between its experiments? The paper is disorganized, making it difficult for a human reader to follow. For example, the paper presents empirical results in the introduction without fully discussing the background and motivation. What is the purpose of presenting it that way? Can the authors also provide a more comprehensive introduction to the limitations of existing work?
  2. How should the improvement from same-model verification be interpreted, and what are its main limitations? It would also be useful to explain cases where the verifier strongly favors an incorrect candidate even when a correct candidate is available. This would help readers understand when self-verification is useful and when its judgments remain unreliable.
  3. Given the closely related work by Gekhman et al. [1], can the authors clarify what principal conclusions this paper summarizes, and what novel contributions this paper clearly makes? Although the paper is analytical and informative, it does not clearly state, guide, or propose how the findings can contribute to the development of future methods for factual answer generation and selection.

References: [1] Gekhman, Zorik, et al. "Inside-out: Hidden factual knowledge in llms." COLM 2025.

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

This paper empirically studies the gap between generating a correct candidate and selecting it as the final answer in factual question answering. It finds that pre-generation internal readouts predict candidate coverage better than selection success, while explicit self-verification generally improves candidate ranking over generation likelihood but yields modest and model-dependent accuracy gains. It further shows that stronger likelihood baselines reduce these gains and that permissive reference matching can obscure them by accepting unresolved answer lists.

Critical strengths

The paper investigates an interesting question: why language models may fail to select a correct answer even when one is present among their generated candidates. It systematically analyzes candidate coverage, ranking, and final selection across multiple models, examining the roles of internal knowledge readouts, explicit self-verification, generative scoring baselines, and correctness criteria.

Critical weaknesses
  1. The paper’s main contribution lies in its empirical observations rather than methodological innovation, but the central findings appear largely intuitive and offer limited new insight. For example, generating a correct candidate does not imply that a selector can identify it, and modest improvements from explicit verification are plausible given the additional computation and a task-specific evaluation prompt. Although these findings merit quantification, the paper does not sufficiently establish what new understanding they provide beyond prior work. The limited gains over stronger likelihood baselines and the lack of clear benefits under matched computational budgets further weaken their practical significance.
  2. The experiments are restricted to short English factual questions involving a limited set of entity types and relations. It remains unclear whether the findings generalize to long-form generation, multi-step reasoning, specialized knowledge domains, or substantially larger models. This narrow scope limits the broader implications of the analysis.
Questions for authors

Beyond the empirical observations, what concrete practical techniques or actionable guidelines does the paper offer, and how are they supported by the results?

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

The paper separates factual recall, candidate coverage, ranking and final selection. It compares likelihood with P(True), then examines how prompting and correctness rules affect verification gains across models.

Critical strengths

1.Fixed candidate pools keep the scorer comparison interpretable. Reporting final selections alongside AUROC shows whether improved ranking actually changes the answer.

2.Stronger likelihood baselines, measured-cost comparisons and blinded annotations address several explanations of the original gain. The paper reports the cases where that gain weakens.

Critical weaknesses

1.The main cross-model comparisons reuse one short-question cohort, with bare-format generation and format filtering. Chat-template rescoring leaves that candidate-generation process unchanged. The primary correctness rule also rejects some outputs with a correct final choice; the later human validation comes from the authors on the same Gemma cohort. The main findings remain closely tied to this evaluation setup.

2.Greedy decoding already performs close to verification plurality on the main Gemma cohort. When likelihood is credited as available during generation, the cost-matched gain has an interval spanning zero; Llama's cost comparison is also inconclusive. The evidence for spending extra inference on verification is much weaker than the ranking result.

Questions for authors

1.Which reported result best supports a verification advantage after both prompt-format and inference-cost differences are addressed?

2.Which conclusion is intended to generalize beyond the shared selection cohort and its correctness rules?

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

Related papers

Discussion (0)

Sign in to comment.