acceptodds
Under review as a conference paper at ICLR 2027

Semantic Uncertainty Quantification Needs Factual Equivalence

29 views

Abstract

Semantic uncertainty quantification for large language models rests on a common template: sample several answers, measure how much they agree, and treat disagreement as uncertainty. We first formalize this template as two separate roles: an operator that compares two answers, and an aggregator that combines all pairwise comparisons into a scalar. Existing methods differ almost entirely in how they aggregate, while taking the operator off the shelf, typically an NLI model or a generic sentence encoder. We show that this reliance on off-the-shelf operators is the primary bottleneck of semantic UQ: they do not accurately measure factual equivalence of multiple answers to the same question. We resolve this with a deliberately simple recipe: a single encoder trained contrastively to isolate the targeted fact, utilizing synthetic data generated by an LLM and dataset both disjoint from all evaluation settings. Integrating the resulting operator into existing methods improves performance on 120 of 126 evaluation settings (95%) spanning 18 model–dataset combinations across language and vision-language models. The best variant reaches 0.76 mean AUROC against 0.68 for the strongest baseline, while replacing the quadratic cross-encoder comparisons of entailment-based operators with one encoder pass per answer. The uniformity of the improvement supports the view that the operator, not the aggregator, is the limiting factor. The same operator also improves single generation token-level estimators: the norm it assigns to each token measures how much that token bears on the answer, and reweighting token log-likelihoods accordingly sharpens the estimate.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty Quantification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RoFvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper decomposes semantic uncertainty quantification into a comparison operator and an aggregator, and argues that the former is the main bottleneck. The authors train a question-conditioned bi-encoder for factual-equivalence comparison and use it as a drop-in replacement for the operators in SE, CAE, KLE, and COS, reporting consistent improvements across LLM and LVLM benchmarks while reducing pairwise comparison cost from O(N2)O(N^2) to O(N)O(N). The learned representation is also used to improve token-level uncertainty estimation

Critical strengths
  1. The operator–aggregator decomposition provides a clear and useful perspective on semantic UQ.

  2. The proposed operator is simple and computationally efficient.

  3. The experimental coverage is broad.

  4. The improvement from replacing the operator is remarkably consistent.

  5. The paper includes several useful ablations and diagnostic experiments.

  6. The token-level analysis is an interesting byproduct of the learned representation. The highest-weighted token falls inside the ground-truth answer span in 72% of the selected TriviaQA examples, compared with 3.9% under random selection. This suggests that the learned representation captures answer-relevant content beyond its use as a pairwise comparison operator.

Critical weaknesses
  1. The training objective does not align with the paper's claims. The authors use correctness labels to construct contrastive learning samples, but answer correctness is not equivalent to semantic equivalence between answers. For example, two incorrect answers may be semantically equivalent or non-equivalent.

  2. The motivating “Entailment Trap” example is not convincing as presented. I reproduced the paper’s example, “What is the capital of France?”. When the question is prepended to both answers, which is standard practice in Semantic Entropy and is also explicitly described in Appendix A.2 of this paper, the bidirectional entailment probability is very high (around or above 95% in my test), rather than the low similarity suggested by the motivating example. Appendix A.2 states that each generation is concatenated with xtextx_{\text{text}} before NLI comparison. Therefore, this example does not appear suitable for demonstrating an intrinsic failure of question-conditioned NLI.

  3. The primary cause of low entailment appears to be that free-form generated answers contain a large amount of additional factual information, which I do not consider an inherent limitation of the NLI operator. For example, two answers may both correctly state that Thriller was released on November 30, 1982, while adding different claims about release dates in other countries. In such cases, low full-response entailment is not necessarily an NLI failure, because the two complete responses are genuinely not factually equivalent. Following the authors’ setup, I tested 1,000 question–answer pairs from NQ-Open. When the model was instructed to generate short or structured answers, the average entailment score for factually equivalent answers increased to above 0.8. This suggests that a substantial portion of the so-called Entailment Trap can be addressed through generation-side control, without the need to retrain the operator.

  4. Most citations in the paper should use \citep{} rather than \citet{}.

Questions for authors
  1. Could the authors clarify the exact NLI input construction and scoring rule used for the motivating “Entailment Trap” example and for Appendix F.3? Was the question prepended to both answers exactly as described in Appendix A.2 and in the original Semantic Entropy implementation?

  2. Does the proposed operator still provide substantial gains when answer length and output format are controlled? Since concise or structured generation substantially improves NLI entailment in my tests, I would like to see a comparison such as:

  • original NLI + free-form generation,
  • proposed operator + free-form generation,
  • original NLI + concise/structured generation,
  • proposed operator + concise/structured generation.
  1. How do the authors define “factual equivalence” in the presence of contradictory ancillary information? If two responses give the same answer to the question but contain conflicting additional claims, should they still be considered semantically equivalent?

  2. Could the authors provide the trained encoder weights and inference code during the review period?

1 comment

Thank you for your response. I successfully reproduced the experimental results using [huggingface.co/hugginganonym/Encode_ICLR27], and integrating it into my own method also led to an improvement in the average AUROC. I have updated my score to 3.

Reviewer_8qjGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper focus on the semantic uncertainty quantification for LLMs. It decomposed the previous method into two components: operator that compares multiple answers, and an aggregator that combines all comparisons into a scalar as the level of uncertainty (confidence score). This work argue that the bottleneck of semantic uncertainty quantification is the operator. The existing operator has some flaws, such as context trap, entitlement trap, autoregressive bias, etc. And propose training an encoder in a contrastive way as the operator. This work leverage six models and six datasets to demonstrate the effectiveness of the proposed encoder.

Critical strengths
  1. Although not proved systematically, I do agree that the limitation of semantic uncertainty quantification for LLMs mostly exists in the operator part. And the bottlenecks listed in 3.2 are valid and causing the suboptimal performance of semantic UQ.

  2. The idea of training an encoder model is straightforward and well-written in the paper. The trained operator can be directly substituted into existing semantic UQ methods without modifying their aggregators.

  3. The experimental results shows the effectiveness of the trained encoder.

Critical weaknesses
  1. The novelty of using contrastive learning to build the encoder is limited. This paper has a gap between motivation and proposed solution. The experimental results only show aggregated scores. The bottlenecks listed in 3.2 are not proved systematically. And the proposed method are not analyzed against those bottlenecks.

  2. Lack of baseline comparison. From a paper writing perspective, the operator in each baseline estimator are not introduced. Furthermore, there are many strong NLI models or embedding models available as the operator (e.g., leading models on the mteb leaderboard), which are not compared with the trained model in this work.

  3. Task assumption issue. The experiments (both training and evaluation) only focus on the QA dataset with short generation. This proposed method is not tested on long-generation tasks.

  4. The issue LLM-as-a-judge usage. This training data is build using Qwen2.5-32B-Instruct and the correctness is assessed using Llama-3.1-70B-Instruct. And the metrics are reported as incorrect generations vs correct ones. It seems that LLM-as-a-judge (Qwen2.5-32B-Instruct) is the upper-bound of the proposed semantic UQ method, while it does not require sampling multiple responses. Therefore, the LLM-as-a-judge should be included as a baseline.

Questions for authors
  1. What is the performance of the operator in each baseline estimator? How does the existing cross-encoder NLI model or strong embedding model compare to the trained encoder in this work?

  2. Can you systematically (not by using a single example) demonstrate the bottlenecks in existing semantic UQ methods? To what extend does the proposed method address the bottlenecks listed in 3.2?

  3. Will this encoder work for UQ on long-generation tasks?

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

The paper formalizes semantic uncertainty quantification (UQ) as U=F(S), explicitly separating a pairwise factual-equivalence operator s from an aggregator F. It argues that the true bottleneck is not the aggregator but the off-the-shelf operator: external sentence encoders suffer a context trap, internal hidden states suffer autoregressive bias, and NLI cross-encoders suffer an entailment trap plus distribution mismatch and O(N2) cost. The authors propose a question-conditioned bi-encoder trained with a contrastive triplet objective using weak supervision from the LLM's own predictive variability: correct generations form positives, same-question incorrect generations form hard negatives. At inference, the frozen encoder produces a symmetric similarity matrix in O(N) forward passes and serves as a drop-in operator for SE, CAE, KLE, and COS. Encoder norms are also used for token-level TE reweighting. Experiments cover 6 models and 6 datasets (18 model-dataset combinations), reporting improvements in 120/126 settings, mean AUROC from 0.68 to 0.76, and roughly 30× lower operator cost than NLI baselines.

Critical strengths
  1. Operator-centric reformulation and diagnosis. The paper reframes semantic UQ as operator plus aggregator and identifies the equivalence operator, not the aggregator, as the bottleneck. The diagnostic desiderata D1–D3 make this reframing principled and likely to shift future research from aggregator tuning to operator design.
  2. Weakly supervised question-conditioned operator with practical drop-in value. The paper provides a reusable, weakly supervised, question-conditioned equivalence operator trained with hard negatives from the LLM's predictive variability, requiring no manual labels. This decouples operator quality from aggregator design and offers drop-in infrastructure for sample-based UQ.
Critical weaknesses
  1. The method only trains on correct–correct positives and correct–incorrect negatives; it never directly validates whether two incorrect-but-equivalent generations cluster together. Since semantic UQ crucially depends on consistent hallucinations collapsing into one cluster, this is a core mechanism gap. B.1 gives only one illustrative example, not a systematic evaluation.
  2. Training uses only Llama-3.1-8B on NQ-OPEN, and all evaluation/diagnostic benchmarks are short factoid QA (HotpotQA, TriviaQA, WebQuestions; ADVQA/OKVQA/VizWiz). The paper’s central claim is about factual equivalence as a general operator, but evidence is confined to short-answer factoid regimes. “Factual equivalence” is not all of semantic equivalence; long-form, multi-answer, ambiguous, and reasoning-heavy generations are not tested.
Questions for authors
  1. Does the paper systematically evaluate whether incorrect-but-equivalent generations cluster together, beyond the single illustrative example in B.1?
  2. Do the reported experiments establish that the operator generalizes beyond short factoid QA and text-only training—specifically to long-form, multi-answer, ambiguous, reasoning-heavy, and visually grounded generations?
  3. Does the paper quantify LLM-judge label noise and its effect on the learned factual-equivalence geometry, e.g., through judge agreement, human audit, or synthetic-noise robustness?

Comments on the paper

2 comments

We thank the Area Chair and the reviewers for their time and feedback, and we are happy to provide more details or experiments if needed. Here is a short summary of our responses for each reviewer.

Reviewer NCvt. We ran three additional experiments:

  • an evaluation on TriviaQA questions, showing that equivalent incorrect answers cluster together (0.903), while different incorrect answers are kept apart (0.448);
  • an analysis by answer length, showing that our gain over NLI grows from +0.04 to +0.09 on longer generations ;
  • a test with three additional LLM judges, where our operator outperforms the strongest baseline in every case (+0.05 to +0.10 AUROC).

Reviewer 8qjG. Several of the concerns are already addressed in the submitted paper, and we pointed to the relevant sections:

  • the diagnostics of each bottleneck (Sec. 3.2, App. F);
  • the baseline operators (Sec. 3.1, App. A.2, E.3).

Following the reviewer's suggestions, we also ran new experiments:

  • a comparison with two leading MTEB embedders more than 20× larger (7–8B vs. 335M parameters), where our encoder remains clearly the best (0.759 AUROC vs. 0.703 and 0.705);
  • an analysis by answer length, showing that the advantage of our operator grows with length, from +0.04 to +0.09 AUROC over NLI on longer generations.

Reviewer RoFv.

  • We ran the requested comparison of free-form vs. concise generation, where our operator remains the best in both cases.
  • We clarified the NLI protocol.
  • For full reproducibility, we shared the code, the data, and the checkpoint on an anonymous Hugging Face page.

We thank the reviewer for successfully reproducing our results, and raising their score.

 

We hope these responses address all the concerns raised, and we will be happy to discuss further if any concerns remain.

Dear Area Chair and Reviewers,

We would like to kindly remind that the Author-Reviewer discussion ends in nearly 24h, and we haven’t had much opportunity so far to discuss this with you. We wanted to let you know that we would be happy to discuss any remaining questions or concerns, or to provide further details on our responses.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.