acceptodds
Under review as a conference paper at ICLR 2027

Self-Brier and Formal Semantic Entropy

Abstract

An LLM may give conflicting answers to the same question while reporting high confidence in each. We introduce two uncertainty scores that summarize its expressed certainty across repeated outputs without reference answers. Self-Brier replaces the observed outcome in the Brier score with an indicator of agreement with the most frequent sampled answer. It averages the squared differences between confidence and this indicator, combining doubt among supporters with confidence among dissenters. Formal Semantic Entropy groups formalized claims by logical equivalence and computes entropy from the group frequencies. An auxiliary model translates responses into formulas over a shared vocabulary. A theorem prover, here the SMT solver Z3, checks equivalence under shared assumptions. The groups depend on which claims the translation retains and how faithfully it represents the response. We study three models on four public reasoning benchmarks. Formal entropy covers FOLIO, ZebraLogic, and GSM-Hard. In a secondary error-detection evaluation, Self-Brier has higher mean AUROC than voting on all four benchmarks, but its gains over confidence remain uncertain. Formal entropy has higher mean AUROC than our binary LLM-equivalence baseline on all three formalized benchmarks; confidence has higher mean AUROC than formal entropy on each.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
uncertainty quantification, large language models, LLM-as-a-judge, selective prediction, self-consistency, semantic entropy, formal methods

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_4af1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes two methods to turn 3 samples of (class, confidence) tuples per question into an uncertainty score, and measures AUROC. Class here is derived obtained by feeding raw answers into an LLM judge (or a formal equivalence checker), and confidence is prompted from the answering model.

Critical strengths

It’s nice that there are 4 datasets.

Critical weaknesses
  1. The proposed methods are not really novel or exciting. Essentially Semantic Entropy applied to a specific way of defining semantic equivalence.
  2. Looking at Figure 6, it looks like the simple “Confidence” baseline (average the confidence statements the model is prompted for) is on par with or outperforms the proposed methods. This is in line with previous works that propose new scores for UQ, where often it does not really beat the baseline.
  3. Some text is quasi-not understandable and reads AI written, often containing undefined words and being very vague. For example the very first sentence: “A confidence report describes a single response” → what even is a confidence report? This word is used twice in the paper and never defined. Also, this whole paper is not about a single response but question-level uncertainties, aggregating 3 responses. Another example: “Let A be the indices with a parsed answer.”. What indices? i? And what is a parsed answer? Do you mean “We use an LLM judge to map raw answers a_i to predefined classes, and A contains all i where this was successful”? No need to respond to these examples in general, but please rewrite the paper so that it is mathematically and logically precise.
  4. The eval does not seem to be fair. “LLM entropy [a baseline method, which is never defined neither in main paper nor appendix] groups responses using pairwise equivalence judgments from Claude Sonnet 4.5, the auxiliary model that also supplies the formalizations. This differs from the evaluated Sonnet 4.6 model.”
  5. The eval uses 100 questions per dataset to generate ROC curves and AUROC, which is way below what is necessary for stable results. Check for example Figure 6 in the appendix, there is no clear trend.
  6. The paper claims “Once answers and confidences are available, it requires no additional model calls.“ but also implicitly assumes there is some LLM judge mechanism that maps answers to categories, and also that there are meaningful categories to be mapped to in the first place. This mechanism is not described, or even pointed out that it exists.
Questions for authors

None.

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

The authors propose two uncertainty metrics for LLMs: Self-Brier, which combines reported confidence with agreement with the modal answer, and Formal Semantic Entropy, which computes entropy over clusters of logically equivalent formalized claims. Experiments with three models on four reasoning benchmarks show that self-brier achieves higher mean error-detection AUROC than voting, while formal semantic entropy improves over a binary LLM-equivalence baseline on the three formalized benchmarks.

Critical strengths
  1. The adaptation of the Brier score from weather forecasting to LLM uncertainty estimation is interesting.
  2. The paper provides detailed experimental settings and prompt templates, making it easier for readers to reproduce the results.
Critical weaknesses
  1. The presentation needs substantial improvement. The paragraph organization and transitions between points are often unclear, making the discussion difficult to follow. In addition, some table captions are placed below the tables, contrary to the ICLR formatting guidelines, which require them to appear above.
  2. The empirical results do not demonstrate the superiority of the proposed Self-Brier and Formal Semantic Entropy metrics over the mean confidence baseline. In fact, they perform worse on most of the selected benchmarks. This leaves the added value of the proposed metrics for uncertainty estimation unclear.
  3. Beyond the empirical results, the paper offers limited new insights. Most findings appear to reiterate observations already reported in prior works and contribute little to advancing research on LLM uncertainty estimation.
  4. The evaluation is insufficiently connected to prior uncertainty estimation research. The selected benchmarks have little overlap with those used in prior work, and relevant baselines such as Avg-Conf are missing [1]. Since Avg-Conf also combines verbalized confidence with answer consistency, a direct comparison with it is particularly important.
  5. Formal Semantic Entropy not only underperforms the mean-confidence baseline but also has narrower application scenarios. It first requires converting a task into Z3 representations, which can be difficult to achieve for many real-world tasks.
Questions for authors
  1. How does changing the confidence values in the few-shot examples affect the reported results?
  2. What reasoning effort settings were used for each model in the experiments? Do the findings hold for stronger models, such as Claude Opus, GPT-6 Astra, or DeepSeek-V4.1-Flash?

[1] Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. Can LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs. In The Twelfth International Conference on Learning Representations, 2024.

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

The paper proposes two ways to measure how uncertain a language model is across repeated answers to one question, without reference answers. The first, Self-Brier, compares each sample's stated confidence with whether it agrees with the most common answer, and rises when supporters of the majority answer sound unsure or when dissenters sound sure. The second, Formal Semantic Entropy, has a helper model translate each answer and some of its reasoning into logic, uses a solver to group translations that say the same thing, and measures how spread out the groups are. Across three models and four reasoning benchmarks, Self-Brier detects wrong answers better than majority voting on average but not clearly better than the model's own confidence. The formal score beats a baseline that groups answers with an LLM judge but stays below confidence everywhere.

Critical strengths
  1. I think Self-Brier is a simple and useful idea. It needs no extra model calls, and the worked examples show clearly how it separates confident agreement from confident dissent when average confidence and the vote are the same.

  2. I appreciate how candid the paper is. It states what the solver does not guarantee, reports results that go against its own methods, and documents prompts, failures and costs in detail.

  3. Formal Semantic Entropy makes the rule for when two answers mean the same thing explicit. Because a solver checks equivalence over a shared vocabulary, we can see why two samples were grouped together, which is often left vague when a model judges meaning.

  4. The evaluation covers four benchmarks and three models with ten samples per question. They report selective prediction alongside AUROC, use paired bootstrap intervals over questions. They add checks such as using five samples instead of ten, removing parts of Self-Brier and controlling for failed translations.

Critical weaknesses
  1. Neither score improves on the plain average of the model's stated confidence. Every Self-Brier interval for this comparison includes zero and the formal score is always below confidence, so I feel the practical case for both scores is not yet made.

  2. The main gain of the formal score is over an LLM baseline whose prompt counts two answers as the same only if they use the same premises, rules and key step, and which defaults to "different" when unsure. This baseline is below chance on ZebraLogic, so I think plain answer entropy is the fairer comparison, and there the gains are small and come without intervals (Table 2).

  3. The formal score depends heavily on the translation. The paper has no measure of translation fidelity, nearly all formulas for one model on GSM-Hard contain only the final answer, and the FOLIO advantage almost disappears when only fully valid translations are kept, though that subset also changes the questions.

  4. The evidence is thin in places. There are 100 retrospectively chosen questions per benchmark, ZebraLogic has eleven scored errors in total, and prompts differ across models, as the authors acknowledge.

Questions for authors
  1. Could the authors report paired intervals for the formal score against answer entropy, which seems the closest baseline that ignores wording?

  2. How does Self-Brier compare with a simple combination of mean confidence and vote fraction, such as their average? If such a combination matched it, I think the case for the squared form would be weaker.

  3. How often does the LLM baseline split samples with the same answer, and would a looser prompt change the comparison?

Suggestions

  1. A small human check of translation fidelity on a sample of formulas might help separate real differences in reasoning from translation noise.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.