- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies uncertainty estimation for LLMs when a question has multiple valid answers. The authors introduce MACE, a benchmark with 12,000 questions from six domains. Each question has 1, 2, 4, or 6 valid answers.
The authors find that when a question has more valid answers, LLMs often give lower confidence scores, even though their answer accuracy increases. They call this problem the concentration–correctness gap. They also show that many uncertainty estimation methods perform worse on questions with multiple valid answers. To address this problem, they propose CARE, a simple method that adjusts confidence scores based on the number of valid answers. Experiments show that CARE improves AUROC compared with the original consistency-based method.
Critical strengths
- The motivation is clear
- The authors introduce a new benchmark with different numbers of valid answers, which is useful for studying uncertainty estimation
- The paper evaluates many existing methods and provides findings about them across different answer cardinalities
Critical weaknesses
- The writing part should be improved. There are many compound sentences connected by punctuation, created terms, and some paragraphs directly go to very technical details without introducing the background or motivation
- The definition and novelty of the concentration–correctness gap need further clarification. The gap is not unique to multi-answer questions. A model can be highly concentrated on an incorrect answer even when only one valid answer exists. This is supported by Table 1&2, but it is not clear how the proposed concentration–correctness gap is formally defined and how it is associated with the ranking capability of an uncertainty method.
- The definitions of query-level and answer-level uncertainty are unclear and potentially misleading. The paper defines query-level uncertainty mainly through the model's output distribution. However, query-level uncertainty can also be estimated before answer generation [1]. The authors should clarify this difference.
- The proposed CARE method lacks theoretical justification and has limited practical value. The method appears to be heuristic, without a clear explanation of why this form should estimate answer correctness. CARE assumes that the true number of valid answers (K) is known. In real-world applications, this information is usually unavailable.
- It is not clear why report the cumulative answer-cardinality
- P(True) is very stable and robust against larger K, which undermines the contribution of this work
- The benchmark only keeps popular facts, which is not aligned with real-world questions.
[1] Query-level uncertainty in large language models. ICLR 2026
Questions for authors
See weaknesses
7 comments
Response to Reviewer sWPX
Response to Weakness 5: Why Evaluate Cumulative Mixtures
Thank you for raising this question. AUROC measures how well confidence ranks correct responses above incorrect ones. In deployment, questions with different numbers of valid answers arrive together, and a system may apply one confidence threshold to decide which responses to accept or reject. Scores must therefore be comparable across answer counts. For example, a correct response to a multi-answer question may receive a lower score than an incorrect response to a single-answer question because its support is spread across valid alternatives. Separate-cardinality AUROC never tests this comparison: it only compares correct and incorrect responses within the same answer count. An estimator can therefore rank well within each count while ranking poorly when the groups are mixed.
Table 3 progressively includes multi-answer questions to assess this comparability. Consis falls from 80.1 AUROC on single-answer questions to 74.6 on the full mixture. When comparing correct responses to questions with six valid answers against incorrect responses to questions with one valid answer, AUROC is only 59.5 (Table 20), showing why cross-count comparisons matter.
We therefore prioritize pooled AUROC, with separate-cardinality results in Appendix Table 18 as complementary diagnostics. We will present this rationale more clearly in Section 5.
Response to Weakness 6: Stability and Discrimination of P(True)
Thank you for raising this point. We would like to clarify that PTrue-Consis-Cand’s relative stability is itself one of our findings (Section 5.3). Our mechanism analysis suggests that separate verification may reduce reliance on generation dominance, making it an important direction for further research. Stability matters, but so does absolute discrimination. Table 3 reports the following AUROC results for LLaMA-3.1-70B-Instruct:
| Estimator | Single-answer AUROC | Full-mixture AUROC |
|---|---|---|
| PTrue-Consis | 68.4 | 63.3 |
| PTrue-Prob | 72.5 | 66.2 |
| PTrue-Prob-Cand | 75.1 | 70.6 |
| PTrue-Consis-Cand | 62.5 | 62.4 |
| Consis | 80.1 | 74.6 |
| CARE (verified counts) | 80.1 | 78.9 |
The nearly stable variant achieves 62.4 pooled AUROC versus 74.6 for uncorrected Consis, while the other three P(True) variants decline. CARE addresses the instability of a stronger estimator: with verified counts, it raises Consis’s pooled AUROC to 78.9 and reduces its decline from single-answer performance. Improving separate verification and correcting strong estimators such as Consis are complementary directions toward confidence that is both stable and discriminative.
Response to Weakness 7: Purpose of Popularity Filtering
Thank you for raising this point. We would like to clarify that MACE provides controlled, closed-book confidence evaluation across answer counts. Studying valid-answer dispersion requires some model support for valid alternatives. Popularity filtering (Section 3.2; Appendix B.3) reduces the risk of unfamiliar entities dominating the evaluation, which would mainly test knowledge gaps or the need for retrieval. Limiting the candidate pool also controls expert verification costs.
Excluding obscure entities does not mean that the remaining questions are easy. Even after filtering, LLaMA-3.1-70B-Instruct achieves only 48.0% single-answer accuracy (Table 2), showing that the benchmark remains challenging. It reflects a practical use case: factual questions answered from the model’s existing knowledge without external retrieval. This design supports controlled diagnosis of confidence reliability in a relevant application setting.
Dear Reviewer sWPX,
Thank you again for your valuable comments on our paper. As the rebuttal period draws to a close, we would appreciate your feedback on whether our responses and additional experiments have adequately addressed your concerns.
If any questions remain, we would be happy to provide further clarification. If our responses have resolved your concerns, we would be grateful if you would reconsider your assessment and score in light of the additional evidence.
Thank you for your time and consideration. We look forward to hearing from you.
Best regards, The Authors
Response to Reviewer sWPX
Response to Weakness 1: Writing and Presentation
Thank you for your suggestions on presentation. We will make the exposition clearer, shorten compound sentences, and streamline terminology.
Response to Weakness 2: Novelty and Definition of the Concentration–Correctness Gap
Thank you for raising this distinction. We would like to clarify that our finding concerns reduced concentration on a correct response because other valid answers share the support. In your single-answer example, probability concentrates on a wrong answer, leading to high confidence in an error. In our case, total valid support remains stable, but its division among correct alternatives lowers confidence in the returned correct answer.
For a returned correct answer, the gap can be formally defined as the difference between total probability mass on all valid answers and the mass on the returned answer, measured over the semantic answer distribution. It captures support allocated to other correct alternatives. Table 2 illustrates this: among correctly answered questions at each answer count, total valid support remains around 90–92%, while support for the returned answer falls from 90.2% to 70.0%.
This gap affects correctness ranking because many estimators derive confidence from concentration through answer likelihood or sampling agreement. Valid-answer dispersion can therefore lower a correct response’s score enough to rank it below an incorrect response. Ranking metrics measure this loss of alignment between confidence and correctness, complementing the distributional evidence for the gap.
We agree that this definition and connection were not clear enough and will clarify them in the revised manuscript.
Response to Weakness 3: Query-Level and Answer-Level Uncertainty
Thank you for pointing out this distinction and reference [1]. Our study follows the established post-generation evaluation setting. Within this setting, our taxonomy distinguishes query-level uncertainty over the sampled output distribution, represented by semantic entropy [2], from answer-level confidence in a particular response, represented by P(True) [3]. We acknowledge that “query-level” may be ambiguous given its use for pre-generation answerability in [1]. We will clarify this terminology and discuss this complementary direction in Related Work.
[1] Query-Level Uncertainty in Large Language Models. ICLR 2026.
[2] Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation[J]. arXiv preprint arXiv:2302.09664, 2023.
[3] Language models (mostly) know what they know[J]. arXiv preprint arXiv:2207.05221, 2022.
Response to Weakness 4: Motivation, Design, and Practical Value of CARE
Thank you for raising this point. We would like to clarify the motivation behind this transformation. As Table 2 shows, when multiple answers are valid, support for a correct answer can decrease because other valid answers share the model’s support. CARE uses answer count to compensate for this potential dilution without identifying which generated answers are correct.
Linear rescaling is intuitive, but valid support is not equally distributed. CARE instead uses a saturating transformation: for a fixed count, it behaves like count multiplication near zero, then smoothly saturates as frequency increases. With verified counts, it preserves single-answer scores and rankings within each fixed count.
Experiments with verified counts isolate the value of cardinality information. Since counts are usually unavailable in practice, we also tried predicting them from model-output features using a regressor (Appendix D.12). All results below are for LLaMA-3.1-70B-Instruct (Tables 3 and 34). Recovery measures the fraction of uncorrected Consis’s observed AUROC loss regained in the full mixture:
CARE outperforms linear rescaling in both settings: it recovers more of the loss with verified counts and remains beneficial with estimated counts, where linear rescaling degrades AUROC. Estimated-count gains remain limited, but the verified-count results demonstrate the value of cardinality information. We identify reliable count estimation directly from the question as future work in Section 6.