acceptodds
Withdrawn from ICLR 2027

ACCOUNTING FOR THE UNSEEN: FRAGILITY-AWARE RECALIBRATION OF SEMANTIC UNCERTAINTY FOR HALLUCINATION DETECTION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Despite their impressive ability to generate natural and coherent text, large language models (LLMs) can produce plausible but fabricated responses. The same prompt, when repeatedly applied, may yield different answers, undermining the confidence one may place in an individual generation. The common approach of detecting such confabulations by observing multiple generations to estimate the uncertainty over their semantic content must necessarily rely on a finite sample of all generations that the model may produce. Such estimates can be fragile when rare or singleton answers dominate observed semantic support. To address this, we propose Semantic Uncertainty Recalibration accounting for Fragility (SURF), a framework that corrects semantic uncertainty estimates using the probability of observed singletons as a proxy for unobserved semantic support. SURF incorporates this fragility proxy into an entropy-based optimization that recalibrates token-sequence probabilities according to their sensitivity, yielding uncertainty estimates that are more robust against finite observations. We also provide an estimate of the number of semantically distinct answers as an actionable signal for when additional observations or human intervention may be warranted. When evaluated across 112 different settings (involving four datasets, seven model variants, and different generation lengths and model quantization levels), SURF shows consistent improvement in AUROC and headroom gain (HRG), and maintains stable AURAC performance, which demonstrates its reliability even under resource constraints.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Large Language Models (LLMs), Hallucination Detection, Semantic Uncertainty

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

The paper introduces a SURF framework for uncertainty estimation in LLM.

Critical strengths
  1. Some ideas like accounting for the unseen probability mass are interesting and require further investigation.
Critical weaknesses
  1. The benchmark methods are outdated - the method is compared to SE variants [4,1].
  2. The evaluation, while utilizing the areas under several different curves is poor, the correctness labels compared against are never specified [2].
  3. The novelty is non existent - the confidence based approaches exist for a while now and are well described in terms of risk estimation as well as the probability mass arguments [3].
  4. The clarity is extremely poor. Some sort of optimization is happening in the paper, but the clarity is lacking and I am not sure that it is described in the main paper.
  5. If it is not evident by this point, the paper fails to address whole lines of work in the field.

References:

  1. Vashurin, R., Goloburda, M., Nakov, P., Shelmanov, A. & Panov, M. CoCoA: A Generalized Approach to Uncertainty Quantification by Integrating Confidence and Consistency of LLM Outputs. Preprint at https://doi.org/10.48550/arXiv.2502.04964 (2025).
  2. Ielanskyi, M., Schweighofer, K., Aichberger, L. & Hochreiter, S. Addressing pitfalls in the evaluation of uncertainty estimation methods for natural language generation. in The fourteenth international conference on learning representations (2026).
  3. Aichberger, L., Schweighofer, K. & Hochreiter, S. Rethinking uncertainty estimation in LLMs: a principled single-sequence measure. in The fourteenth international conference on learning representations (2026).
  4. Fadeeva, E. et al. LM-Polygraph: Uncertainty Estimation for Language Models. Preprint at https://doi.org/10.48550/arXiv.2311.07383 (2023).
Questions for authors

see weaknesses

2 comments
AuthorsOpenReview ↗

Clarifying SURF’s Novelty, Baseline Relevance, and Evaluation Methodology (part1)

We thank the reviewer for their assessment and for highlighting recent developments in uncertainty estimation. We respectfully disagree with the characterization that SURF lacks novelty or that its methodological contribution is unsound. We address the concerns individually below.

W1. Outdated Baselines and Comparison with Recent Methods

We respectfully disagree with the characterization that our benchmarks are outdated. Our evaluation includes Shannon semantic entropy (Farquhar et al., Nature 2024), Rényi semantic entropy and sensitivity-aware recalibration (Vipulanandan et al., ICLR 2026), alongside naive entropy, discrete semantic entropy, embedding regression, and p(True)p(\mathrm{True}). These are established and recent approaches directly relevant to the problem investigated in this paper.

Furthermore, the reviewer cites LM-Polygraph [4], originally published at EMNLP 2023, with a subsequent benchmarking study published in TACL 2025. While valuable, LM-Polygraph is primarily a framework for implementing and evaluating uncertainty estimation methods, rather than a competing formulation of finite-sample semantic fragility.

We acknowledge that CoCoA (NeurIPS 2025) and G-NLL (ICLR 2026) are relevant recent developments. However, neither addresses the specific methodological problem considered in SURF.

CoCoA (NeurIPS 2025) combines model confidence with semantic consistency through a minimum Bayes risk formulation. Importantly, its authors explicitly acknowledge that consistency estimates obtained from limited samples may fail to capture the full diversity of model outputs. They further identify task-dependent semantic similarity as a limitation, particularly when subtle reasoning errors are not captured by generic similarity metrics. SURF directly investigates a related but distinct issue: the fragility of the observed semantic distribution under incomplete sampling. Rather than relying solely on observed consistency, SURF introduces singleton semantic mass as an explicit signal of finite-sample support instability and incorporates it into probability recalibration.

G-NLL (ICLR 2026) proposes a computationally efficient single-sequence uncertainty measure based on greedy-decoding likelihood. Although theoretically motivated, G-NLL operates at the sequence-probability level and does not explicitly incorporate semantic equivalence or unobserved semantic clusters. Indeed, its authors identify the incorporation of semantic information as an additional direction requiring investigation. Moreover, greedy decoding only approximates the globally most probable sequence, and the method's uncertainty estimates depend on this approximation. Therefore, G-NLL addresses a fundamentally different objective from SURF's finite-sample semantic uncertainty recalibration.

These distinctions are important: combining confidence and consistency, estimating uncertainty from a single sequence, and accounting for incomplete semantic coverage are not equivalent methodological contributions.

Our baseline selection was designed to isolate the incremental effect of fragility beyond established semantic uncertainty formulations. This is supported by our component ablation across 112 paired evaluation settings, where incorporating fragility improves mean AUROC and combining fragility with sequence-probability sensitivity further improves it.

We agree that including CoCoA and G-NLL would broaden the empirical comparison. However, we respectfully maintain that the absence of these additional comparisons does not justify dismissing the existing benchmarks as outdated or concluding that SURF lacks a distinct methodological contribution.

W2. Correctness Labels and Evaluation Methodology

We appreciate the reference to Ielanskyi et al. (ICLR 2026), which identifies important limitations of approximate correctness functions in uncertainty estimation.

Our evaluation distinguishes between the answer being assessed and the generations used to estimate uncertainty: for each question, one response is sampled at temperature 0.1 for correctness evaluation, while ten additional responses are sampled at temperature 1.0 to characterize semantic variability. The same evaluation setting is used for paired comparisons across methods.

We recognize that the manuscript should explicitly state the correctness-labeling function and associated decision criteria, rather than describe only the evaluation metrics. We will clarify this aspect of the experimental protocol and acknowledge the sensitivity to alternative correctness functions emphasized by the cited work.

The results are evaluated across four datasets, seven model variants, and multiple quantization and prompting configurations, producing 112 paired evaluation settings. These comparisons provide evidence of relative performance under the employed evaluation protocol, although they do not eliminate the need for robustness checks against alternative correctness criteria.

AuthorsOpenReview ↗

Clarifying SURF’s Novelty, Baseline Relevance, and Evaluation Methodology (part1)

W3. Novelty of Fragility-Aware Semantic Uncertainty

We respectfully disagree that SURF merely reintroduces existing confidence-based or probability-mass approaches.

The central contribution is not the introduction of confidence scores, semantic consistency, or the classical concept of missing probability mass. Instead, it is the incorporation of finite-sample semantic support fragility into sequence-probability recalibration.

Specifically, SURF characterizes fragility through the proportion of observed semantic probability mass concentrated in singleton clusters, Fmass=P1/PobsF_{\mathrm{mass}}=P_1/P_{\mathrm{obs}}. This quantity acts as an empirical indicator of incomplete semantic coverage and modulates an entropy-regularized optimization that also incorporates local sequence-probability sensitivity.

The cited works address distinct questions:

  • CoCoA [1] integrates model confidence and semantic consistency within a minimum Bayes risk framework, but does not introduce singleton-occupancy-based fragility recalibration.
  • G-NLL [3] develops a principled single-sequence uncertainty measure based on greedy-decoding likelihood, without explicitly modeling incomplete semantic support under repeated sampling.
  • LM-Polygraph [4] provides a comprehensive benchmarking framework rather than the specific finite-coverage correction proposed here.

We do not claim to recover the exact probability mass or identities of unseen semantic clusters. Rather, SURF uses observed singleton mass as a proxy for the instability associated with incomplete sampling.

Our component ablation further supports this distinction: mean AUROC increases from 0.7695 for Rényi semantic entropy to 0.7775 when fragility is incorporated alone, and to 0.7784 when fragility and sequence-probability sensitivity are combined. These results indicate that the fragility component contributes beyond the entropy-only formulation, although its incremental benefit is modest.

W4. Clarity of the Optimization

Contrary to the reviewer's concern that the optimization may not be described in the main paper, the formulation is explicitly presented in the subsection "Objective Function" within Section 3.

The formulation recalibrates each observed token-sequence probability by optimizing three components: (i) a quadratic Rényi entropy term, (ii) a KL-divergence fidelity penalty weighted by sequence-probability sensitivity, and (iii) a fragility-dependent penalty determined by singleton semantic mass.

The parameter λ\lambda controls fidelity to the original sequence probability, while γ\gamma controls the strength of the fragility correction. The recalibrated probabilities are then used to compute semantic-level uncertainty.

Closing Remarks

We appreciate the reviewer's emphasis on rigorous evaluation and engagement with recent literature. Nevertheless, we believe it is important to distinguish the need for broader baseline comparisons and clearer reporting from a demonstration that the proposed method lacks novelty or is fundamentally unsound.

SURF addresses a specific limitation of finite-sample semantic uncertainty estimation: the potential instability of conclusions drawn from incompletely observed semantic response distributions. Our formulation and component ablations provide a concrete contribution toward addressing this limitation.

We hope these clarifications help the reviewer reconsider the methodological contribution of SURF.

Reviewer_vi8xOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Semantic Uncertainty Recalibration accounting for Fragility (SURF), which extends sensitivity-aware Rényi semantic entropy by introducing an occupancy-based measure with the intention of capturing uncertainty arising from unobserved semantic answers. The authors argue that when an LLM is sampled only a small number of times, the observed semantic clusters may represent an incomplete picture of the response distribution. Consequently, conventional semantic uncertainty estimates may be sensitive to the composition of the sampled answers. The method they propose is combining a fragility measure with token-sequence probability sensitivity via an entropy-regularized optimization. They evaluate the method on four QA datasets, seven model variants, and several quantization levels, reporting improvements in hallucination-detection AUROC. I find the problem motivating this research interesting and the experimental effort significant. However, I have reservations about considering this paper to be presented at ICLR, as I see major problems in the mathematical underpinning of the method as described below.

Critical strengths

I think the paper is strong from several aspects: (1) The paper correctly identifies the deficiency of semantic entropy estimated from a finite set of generated responses. This issue introduces sampling variability, especially under heavy-tailed answer distributions with many low-probability semantic alternatives. The general entropy estimate over observed clusters may underestimate or mischaracterize uncertainty when meaningful probability mass remains unobserved. This is an important problem for hallucination detection. (2) Also connecting semantic uncertainty estimation with the statistical quality of the samples is conceptually interesting. (3) The authors present extensive experimental evaluation of their approach. (4) I think the appendices include useful statistical diagnostics.

Critical weaknesses
  1. My main concern is the mathematical accuracy of Eq. 5. If we consider Eq. 4 and 5 together, then F_mass is a well-defined population quantity, but from what I understand, it is not directly observable from a finite collection of model generations. If I consider what the authors describe in Appendix F2, they approximate unknown population frequencies using a Zipf-regularized empirical histogram for real datasets, then the approach raises several questions: (1) how is the Zipf exponent estimated? How sensitive is the method to model misspecification? Does Zipf regularization impose the very heavy-tailed structure that the proposed fragility measure is intended to detect? There is also a more fundamental distinction between probability-weighted singleton mass and the observable count-based quantity. These two statistics can behave very differently. I expect authors to define F_mass and F^*_mass as the implemented estimator and consider all the statistical implications of differentiating between these two.

  2. Eq. (6), as the main SURF optimization problem, optimizes each adjusted token-sequence probability separately. From what I understand, going through your equations, the quadratic Rényi entropy term is maximized at q=1/2, and the KL term encourages proximity to the original probability p. But the final term -\gamma F_{mass}.q is a linear penalty on the adjusted probability. For positive \gamma, it always favours smaller q, regardless of whether the original probability is above or below 1/2. Then, the correction is not generally equivalent to entropy maximization or a symmetric reduction in overconfidence. I think the authors should make it clear why this particular directional penalty is appropriate.

  3. Equally important to what I said in (2), independently adjusted sequence probabilities need not sum to one. The paper subsequently normalizes semantic-cluster probabilities, but renormalization changes the effective optimization problem. it's not clear to me why independent probability adjustment followed by normalization is theoretically appropriate.

  4. I also see a discrepancy between Algo 4 and Eq. 6. Eq. 6 includes -\gamma F_{mass}.q. Surprisingly, Algo 4 in Appendix E presents a different optimization; it contains the entropy and KL-divergence terms but ignores the fragility correction!

  5. There are more mathematical discrepancies and weak justifications across the paper, including in Eq. 8, Lemmas 1 and 2.

  6. There are major issues and questions with the experimental section; I'll ask in the question section.

Questions for authors
  1. Can the authors establish any bias, consistency, or risk-improvement guarantees for the adjusted missing-mass and unseen-species estimators? In particular, how is the Good-Toulmin count transformed into a valid Bernoulli KL reference probability?

  2. Why are simple singleton-count and Good–Turing scores not included directly as hallucination-detection baselines? Why are KLE, Semantic Density, SNNE, and recent hybrid semantic uncertainty methods excluded?

  3. How did you measure finite-sample robustness? Can you repeat uncertainty estimation using independent generation sets of different sizes and quantify uncertainty-score variance and stability? Does SURF genuinely provide a more accurate estimate of semantic uncertainty, independently of AUROC?

  4. How does the choice of temperature 1 for uncertainty estimation affect the measured correctness of responses sampled at temperature 0.1? How sensitive is fragility to changes in the semantic clustering model and cluster assignment order?

  5. Can the authors establish any bias, consistency, or risk-improvement guarantees for the adjusted missing-mass and unseen-species estimators? In particular, how is the Good-Toulmin count transformed into a valid Bernoulli KL reference probability?

  6. Why are simple singleton-count and Good–Turing scores not included directly as hallucination-detection baselines? Why are KLE, Semantic Density, SNNE, and recent hybrid semantic uncertainty methods excluded?

  7. How did you measure finite-sample robustness? Can you repeat uncertainty estimation using independent generation sets of different sizes and quantify uncertainty-score variance and stability? Does SURF genuinely provide a more accurate estimate of semantic uncertainty, independently of AUROC?

  8. How does the choice of temperature 1 for uncertainty estimation affect the measured correctness of responses sampled at temperature 0.1? How sensitive is fragility to changes in the semantic clustering model and cluster assignment order?

4 comments
AuthorsOpenReview ↗

Addressing the weaknesses identified by the reviewer

We thank the reviewer for the detailed technical assessment. Here we clarify the statistical interpretation of fragility, the optimization formulation, and the relationship between our theoretical and empirical results.

W1. Observability of (F_{\mathrm{mass}}) and Zipf Regularization

Equation (5) defines (Fmass=P1/PobsF_{\mathrm{mass}}=P_1/P_{\mathrm{obs}}), where the cluster masses involve known population probabilities. Therefore, this quantity can generally be evaluated exactly from finite generations. Its purpose is to characterize the probability-weighted fragility of the observed semantic support.

As mentioned in the paper for implementation, the corresponding estimator could be explicitly distinguished:

F^mass=∑ℓ:N(ℓ)=1p^ℓ∑ℓ:N(ℓ)>0p^ℓ\widehat{F}_{\mathrm{mass}} = \frac{\sum_{\ell:N(\ell)=1}\widehat{p}_{\ell}}{\sum_{\ell:N(\ell)>0}\widehat{p}_{\ell}}

Regarding Zipf regularization in Appendix F.2, the real-world NLP text corpora do not provide ground-truth probability distributions. Therefore, we adopt a conventional Zipf-based approximation to assign probability mass to observed elements, consistent with the widely documented heavy-tailed frequency patterns of natural-language corpora. Importantly, this assumption is restricted to the supplementary experiments validating missing-mass and unseen-species estimation and is not employed in SURF's hallucination-detection experiments, where token-sequence probabilities are directly available from the LLM. Thus, concerns regarding Zipf model misspecification do not affect the main SURF framework or its reported hallucination-detection results.

W2. Directional Fragility Penalty in Eq. (6)

The directional penalty is intentional and serves a different purpose from the Rényi entropy and KL terms.

When fragility is high, a large proportion of the observed semantic mass belongs to singleton clusters. This suggests that the sampled responses may not adequately represent the full range of possible answers. Therefore, we introduce a penalty that discourages assigning excessive weight to the observed responses when semantic coverage is uncertain.

We agree that this penalty favors smaller adjusted probabilities and is not a symmetric entropy-maximization term. Rather, it is a fragility-dependent regularization term designed to account for the potential unreliability of the observed sample.

We will clarify this motivation in the manuscript and explicitly distinguish the directional correction from entropy maximization.

W3. Independent Probability Adjustment and Normalization

In SURF, the individual optimization steps are intended to produce recalibrated sequence weights, rather than a probability distribution that must independently sum to one.

This follows the same general procedure used in semantic entropy estimation, where probabilities associated with observed generations are aggregated into semantic clusters and subsequently normalized to obtain a valid distribution.

SURF extends this procedure by first adjusting the observed sequence weights according to their local sensitivity and the fragility of the sampled semantic support. These adjusted weights are then aggregated and normalized at the semantic-cluster level before computing its entropy (i.e Rényi semantic).

We will clarify that Eq. (6) defines independent weight recalibration, while the subsequent normalization constructs the semantic probability distribution used for uncertainty estimation.

W4. Discrepancy Between Algorithm 4 and Eq. (6)

Algorithm 4 in Appendix E presents the sensitivity-aware probability recalibration procedure without explicitly incorporating the additional fragility correction introduced in SURF.

Equation (6) extends this procedure by introducing the term (-\gamma F_{\mathrm{mass}}q), which penalizes adjusted sequence weights according to the fragility of the observed semantic distribution.

We acknowledge that Algorithm 4 should reflect the complete SURF formulation rather than the earlier sensitivity-only objective. We will revise the pseudocode to include the fragility term, ensure consistency in the KL-divergence weighting, and explicitly describe the subsequent semantic-cluster aggregation and normalization.

This is a discrepancy between the presented formulations that requires correction, and we appreciate the reviewer bringing it to our attention.

W5. Mathematical Discrepancies and Justifications

We would appreciate it if the reviewer could elaborate on the specific mathematical discrepancies or insufficient justifications identified in Eq. (8) and Lemmas 1 and 2. In particular, it would be helpful to understand which assumptions, derivations, or conclusions are considered incorrect or inadequately supported.

AuthorsOpenReview ↗

Addressing the weaknesses raised by the reviewer

** Q1. Statistical Guarantees and Good–Toulmin Normalization **

We thank the reviewer for this question. We would like to clarify that the missing-mass and unseen-species estimators are auxiliary contributions evaluated in the supplementary material and are not required for SURF's hallucination-detection procedure.

Regarding theoretical properties, Appendix A establishes the strict concavity and uniqueness of the entropy–KL optimization and derives the limiting behavior of its solution. We additionally investigate the bias and variability of fragility through multinomial and Poissonized sampling approximations.

While these results establish properties of the optimization and fragility measure, we do not claim universal unbiasedness, consistency, or risk-improvement guarantees for the adjusted estimators. Instead, their practical performance is evaluated empirically across uniform, Zipf, Dirichlet, and head–tail distributions, as well as real-world text corpora. The supplementary results include comparisons against classical estimators using estimation error and Upper Confidence Risk (UCR).

Regarding Good–Toulmin, we agree that its output estimates an expected count rather than a probability. Consequently, normalization and clipping are required before using it as the reference parameter of a Bernoulli KL objective.

Q2. Missing Baselines: Singleton Counts, Good–Turing, KLE, Semantic Density, SNNE

We respectfully believe the reviewer may be conflating the objectives of classical unseen-species estimation with hallucination detection.

Good–Turing and related singleton-based estimators were originally developed to estimate missing probability mass and unseen species from finite observations, rather than to detect hallucinations in LLM-generated responses. Accordingly, we evaluate these methods in our supplementary missing-mass and unseen-species experiments, where they serve as directly relevant baselines.

For hallucination detection, our objective is different: to investigate whether incorporating finite-sample fragility improves established semantic uncertainty estimation methods. Our comparisons include Shannon semantic entropy , discrete semantic entropy (Nature 2024), Rényi semantic entropy, and importantly, sensitivity-aware Rényi semantic entropy (SE_RU), a recent ICLR 2026 method, alongside probability-based baselines.

Regarding KLE, Semantic Density, SNNE, and recent hybrid approaches, we recognize their relevance. Our baseline selection was intended to isolate the contribution of fragility relative to established and recent semantic-entropy formulations, rather than provide an exhaustive comparison against every uncertainty estimation method.

Q3. Finite-Sample Robustness and Accuracy of Semantic Uncertainty

We thank the reviewer for this question. Finite-sample robustness is already investigated in the supplementary material under "Paired Statistical and Resampling Analysis."

Specifically, we conduct (i) a paired bootstrap analysis with 1,000 replicates per evaluation setting to quantify uncertainty in AUROC differences, and (ii) a leave-one-generation-out analysis in which one of the ten sampled generations is removed, the clustering-dependent uncertainty scores are recomputed, and the resulting AUROC variation is examined across all 112 settings.

These experiments provide evidence of the stability of SURF's hallucination-detection performance under changes to the evaluation examples and observed generation set.

Importantly, SURF is designed to improve hallucination discrimination by accounting for incomplete semantic coverage, rather than to recover the exact entropy of the underlying semantic distribution. Accordingly, our main evaluation emphasizes AUROC, AURAC, RAC, and HRG.

Q4. Sampling Temperature and Semantic Clustering Robustness

We thank the reviewer. Our protocol was designed for an apples-to-apples comparison with Shannon semantic entropy (Farquhar et al., Nature 2024) and sensitivity-aware Rényi semantic entropy (Vipulanandan et al., ICLR 2026).

We evaluate one response sampled at temperature 0.1 and use ten additional generations at temperature 1.0 to estimate semantic variability, consistent with the compared methods. All methods share the same responses, correctness labels, and clustering procedure, isolating differences in uncertainty estimation.

We use DeBERTa-v2-xlarge-MNLI with bidirectional entailment and hard-argmax decisions. We acknowledge that temperature, entailment-model choice, and cluster assignment order can affect fragility. Our objective is to isolate the contribution of fragility-aware recalibration under matched conditions; robustness across alternative protocols remains future work.

Q5-Q8 We note that Questions 5–8 appear to be identical to Questions 1–4. To avoid repetition, we kindly refer the reviewer to our detailed responses to Q1–Q4 above, where these concerns have been addressed individually

I appreciate the authors for taking the time to provide a detailed rebuttal and, more importantly, acknowledging several inconsistencies in the manuscript. While the response clarifies some of the implementation choices and the intended role of fragility, my principal mathematical and methodological concerns remain unresolved.

  1. The authors argue that F_{mass} can be evaluated exactly from finite generations because token-sequence probabilities are available. However, the population semantic-cluster probabilities defined in Eqs. (4)–(5) require summation over all possible sequences belonging to each cluster, including unobserved sequences. The normalized probability weights assigned to sampled generations are therefore not generally equivalent to these population probabilities. The authors should explicitly define the implemented estimator and justify its relationship to F_{mass}).

  2. I appreciate the clarification that the directional penalty in Eq. (6) is intentional. Nevertheless, this does not resolve the concern that independent weight reduction followed by normalization can increase, rather than decrease, the concentration of the semantic distribution. An empirical improvement in AUROC alone does not establish that the resulting distribution is a better estimate of semantic uncertainty.

  3. The strict concavity and limiting results in Appendix A concern the binary entropy–KL objectives for auxiliary estimators rather than the full SURF objective. Moreover, the binary entropy–KL optimization has the explicit solution which systematically moves the reference estimator toward 1/2. This does not in itself establish improved estimation accuracy or risk. Finally, the bootstrap and leave-one-generation-out experiments are useful but do not establish stability across independent generation sets, which is central to the claimed finite-sample robustness.

AuthorsOpenReview ↗

SURF only needs the empirical, observed-support statistic

We thank the reviewer for these observations and clarify three important points.

W1. Clarification of Observed Semantic Mass in Eqs. (4)–(5)

We respectfully disagree with the reviewer's interpretation that SURF requires recovering the probability mass of unobserved semantic clusters.

Eqs. (4)–(5) characterize fragility over observed semantic support. Specifically, the occupancy conditions (N_\ell>0) and (N_\ell=1) explicitly restrict the calculation to observed clusters and singleton clusters, respectively. Unobserved clusters do not enter the fragility ratio.

In practice, SURF computes this statistic using the normalized token-sequence likelihoods of the sampled generations, aggregated by their observed semantic clusters. Neither the identities nor the probability mass of unobserved generations are required.

As explicitly stated in the manuscript, SURF does not claim to recover the true missing mass. Instead, it uses the relative mass of observed singleton clusters as a proxy for incomplete semantic coverage.

W2. Mathematical Justification of Fragility-Aware Recalibration

We thank the reviewer for the additional mathematical observations. We agree that independent probability recalibration followed by normalization does not necessarily reduce semantic concentration. However, we would like to emphasize that the proposed optimization has well-defined mathematical properties that establish the behavior of the fragility correction.

Let qq denote the adjusted token-sequence weight, pp its original probability, and a=λ/(1+UQ(p))>0a=\lambda/(1+\mathrm{UQ}(p))>0. Equation (6) can be written as

J(q)=−log⁡(q2+(1−q)2)−aDKL(q∥p)−γFmassq.J(q)=-\log(q^2+(1-q)^2)-aD_{\mathrm{KL}}(q\Vert p)-\gamma F_{\mathrm{mass}}q.

1. Strict concavity and uniqueness

The second derivative is

-\frac{8q(1-q)}{[q^2+(1-q)^2]^2} -\frac{a}{q(1-q)}<0.$$ For $0<q<1$, $0<p<1$, and $a>0$, both terms are strictly negative. Therefore, the objective is strictly concave and admits a unique maximizer. **2. Monotonic dependence on fragility** At the optimum, $J'(q^*)=0$. Applying implicit differentiation with respect to fragility, holding $p$ and $a$ fixed, yields $$J''(q^\star)\frac{dq^\star}{dF}-\gamma=0.$$ Hence, $$\frac{dq^\star}{dF}=\frac{\gamma}{J''(q^\star)}<0,\qquad \gamma>0.$$ Thus, increasing fragility monotonically decreases the optimal unnormalized sequence weight. These results establish three important distinctions: - **Mathematical guarantee:** The recalibration objective is strictly concave, has a unique optimizer, and produces a monotonic response to increasing fragility. - **Scope of the guarantee:** These properties concern individual unnormalized sequence weights. We do not claim that subsequent semantic-cluster normalization necessarily decreases concentration or increases entropy. - **Empirical objective:** SURF aims to improve hallucination discrimination through fragility-aware recalibration, not to guarantee recovery of the true population semantic distribution. Its effectiveness is therefore evaluated through hallucination-detection performance. We believe these distinctions clarify the theoretical properties of Eq. (6) and the scope of SURF's empirical claims. **W3. Auxiliary Estimator Guarantees and New Finite-Sample Robustness Results** We respectfully emphasize that the binary entropy–KL lemmas in Appendix A concern supplementary missing-mass and unseen-species estimators, not the main SURF hallucination-detection objective. These auxiliary estimators are not required to compute SURF's hallucination scores. Their theoretical guarantees should therefore be distinguished from the empirical claims of the main method. **More importantly, we have conducted additional finite-sample robustness experiments to directly address the reviewer's concerns.** We performed 100 repeated evaluations at each generation count across all 112 model/dataset settings, with 250 questions per setting. | Generations (N) | Mean AUROC | Mean within-setting SD | |---|---|---| | 5 | 0.74289 | 0.01606 | | 6 | 0.75396 | 0.01330 | | 7 | 0.76194 | 0.01113 | | 8 | 0.76799 | 0.00884 | | 9 | 0.77318 | 0.00621 | | 10 (in paper) | 0.77840 | 0.00458 | These results reveal two important trends: 1. **Improved detection performance:** Mean AUROC increases consistently with the number of generations, from 0.74289 at N=5 to 0.77318 at N=9, approaching the originally reported 0.7784 at N=10. 2. **Reduced variability:** Mean within-setting AUROC SD decreases monotonically from 0.01606 to 0.00621, representing approximately a 61% reduction between N=5 and N=9. These additional experiments provide direct empirical evidence that **SURF's hallucination-detection performance becomes more stable as the number of observed generations increases**, consistent with its finite-sampling motivation. We have now incorporated in Appendix C.1 these new results and the corresponding analysis into the revised manuscript.
Reviewer_ydA2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SURF addresses finite-sampling limitations in semantic-entropy-based hallucination detection. The SE methods considered in the paper primarily characterize observed generations without explicitly accounting for incomplete semantic coverage. The authors use the singleton-cluster probability-mass fraction as a coverage proxy and combine it with token-sequence sensitivity to recalibrate token-sequence probabilities before computing semantic uncertainty. Experiments span 112 paired settings across four QA datasets, seven model variants, and three precision levels.

Critical strengths
  1. Clear conceptual distinction. Separating “local probability sensitivity” from “fragility due to incomplete coverage” is a meaningful contribution to the semantic uncertainty literature. The paper clearly explains why sensitivity measured on observed generations does not, by itself, characterize incomplete semantic coverage.
  2. Thorough empirical grid. The 112 evaluation settings cover a broad range of model, precision, and prompting configurations. The component ablation helps assess the contributions of the two correction signals.
Critical weaknesses
  1. Complementarity of UQ and Fmass requires further analysis. Eq. 6 combines two correction terms, but the paper does not report their empirical correlation. The component ablations provide useful evidence, but the full formulation improves mean AUROC over the fragility-only variant by only 0.0009. A paired statistical comparison would help establish whether the additional UQ correction provides a meaningful benefit.
  2. Modest practical gains with uncertain setting-specific benefits. The headline improvement is +0.0089 AUROC over Rényi SE (0.7784 vs. 0.7695). Per-setting bootstrap intervals cross zero in 104/112 settings against Shannon SE and 99/112 against Rényi SE. Although the aggregate results favor SURF, its benefit in most individual deployment settings remains uncertain. For a method proposed for practical hallucination detection, the lack of clear guidance on when to prefer SURF over simpler baselines is a limitation.
Questions for authors
  1. Direct validation of the fragility signal. Can you show, on held-out data, that questions with a higher share of singleton semantic clusters are actually more likely to produce incorrect answers, controlling for question difficulty or model? This would directly test the core assumption linking observed singleton mass to hallucination risk.
  2. Relation between the two correction signals. What is the empirical correlation between the token-sequence sensitivity score and the singleton-mass fragility measure across your 112 settings? If they are strongly correlated, does the full two-term correction actually outperform a correction based on a single combined signal?
  3. Total inference cost. What is the end-to-end wall-clock cost of SURF—including the ten additional samples per question and the semantic clustering step—relative to Shannon semantic entropy with the same sampling budget, and relative to the single-call verifier baseline? For a gain of roughly one AUROC point, this tradeoff matters.
  4. Robustness to clustering choices. How do the singleton-mass estimate and downstream AUROC change if a different entailment model, or a probability-threshold rather than hard-argmax clustering rule, is used? Since the entire fragility signal is computed over clusters, this is a key robustness check.
4 comments

We thank the reviewer for the thoughtful assessment and for recognizing SURF's conceptual distinction between local probability sensitivity and incomplete semantic coverage, as well as the breadth of our empirical evaluation. We address the concerns below.

W1. Complementarity of UQ and Fragility

We agree that the empirical relationship between the two signals deserves further analysis. However, their conceptual distinction is fundamental: UQ characterizes local sensitivity of observed token-sequence probabilities, whereas FmassF_{\mathrm{mass}} characterizes the fragility of the sampled semantic support. Stable sequence probabilities do not imply complete semantic coverage, and conversely, well-covered semantic support does not guarantee locally stable sequence probabilities.

Our component ablation provides evidence of their individual contributions. Relative to Rényi SE (0.7695 AUROC), UQ alone improves performance to 0.7732 (+0.0037), while fragility alone reaches 0.7775 (+0.0080). Combining both yields 0.7784 (+0.0089).

We acknowledge that the additional +0.0009 improvement over fragility-only is small and does not, by itself, establish statistical significance. Crucially, SURF's primary contribution is the fragility correction, which accounts for most of the improvement. The full formulation preserves the complementary UQ mechanism rather than claiming that its incremental contribution is uniformly substantial.

We agree that reporting Spearman correlation between question-level UQ and FmassF_{\mathrm{mass}}, together with a paired comparison of full SURF against fragility-only, would further establish their empirical relationship. Correlation alone, however, would not establish redundancy, since the two signals enter the recalibration objective through different terms.

W2. Practical Gains and Statistical Significance

We acknowledge that the absolute AUROC improvements are modest. However, the results demonstrate a consistent directional advantage across a broad range of evaluation settings, rather than relying on isolated favorable configurations.

Specifically, SURF improves over Shannon SE in 99/112 settings and over Rényi SE in 98/112 settings. The aggregate paired comparisons yield improvements of +0.0069 and +0.0089 AUROC, respectively, with Wilcoxon signed-rank p=1.5×10−12p=1.5\times10^{-12} and p=2.2×10−13p=2.2\times10^{-13}.

We emphasize that confidence intervals crossing zero in individual settings do not establish the absence of an effect; they indicate insufficient evidence to establish a benefit in those particular settings. At the same time, aggregate statistical significance should not be interpreted as guaranteeing improvements in every deployment configuration.

The practical motivation for SURF is not merely maximizing average AUROC. It is to correct a specific deficiency of finite-sample semantic uncertainty estimates: potentially unreliable estimates when observations are dominated by rare semantic responses.

Consequently, SURF is most relevant when multiple generations are already collected for semantic uncertainty estimation and incomplete semantic support is a concern. For applications requiring minimal inference latency or operating with a single generation, simpler methods may be preferable. We will make this distinction clearer rather than suggesting universal superiority.

Q1. Direct Validation of the Fragility Signal

We appreciate this suggestion, but emphasize an important distinction: fragility measures incomplete semantic coverage, not factual incorrectness directly.

A high singleton-mass fraction indicates that the sampled semantic distribution may be unstable under additional observations. It does not necessarily imply that a particular response is incorrect. Rather, the hypothesis underlying SURF is that such instability contains useful information for improving uncertainty-based discrimination.

Our component ablation supports this hypothesis indirectly: introducing fragility without UQ improves AUROC from 0.7695 to 0.7775. Nevertheless, we agree that a held-out analysis of error rates across fragility quantiles, controlling for model, dataset, and baseline uncertainty, would provide more direct evidence of its predictive relationship with correctness.

Q2. Relationship Between the Correction Signals

As discussed in W1, the ablations establish that both signals contribute relative to the entropy-only baseline, but do not establish their statistical independence.

The most informative analysis would compare question-level UQ and (F_{\mathrm{mass}}) through their correlation and conditional predictive contributions, alongside paired AUROC differences between fragility-only and full SURF.

Q3. End-to-End Inference Cost

We have provided the details cost for hamiltonian construction and QTN based UQ calculation in the Supplementary, while the main text states all evaluations were conducted on a workstation equipped with 64 GB RAM and an NVIDIA A6000 GPU.

To reiterate timings : the E2E inference cost is 6-10ms per query post a one time hamiltonian constriction, which takes 45–60 seconds (CPU processing)

Q4. Robustness to Semantic Clustering

We agree that clustering quality is important because (F_{\mathrm{mass}}) depends on semantic cluster occupancy.

Our experiments employ a fixed DeBERTa-v2-xlarge-MNLI model with bidirectional entailment and hard-argmax decisions, consistently applied across all evaluated methods. This controls clustering-related variation in the paired comparisons.

Nevertheless, consistent clustering does not establish robustness to alternative clustering rules. We agree that evaluating alternative entailment models and probability-threshold-based clustering would help quantify the sensitivity of singleton occupancy and downstream AUROC.

Importantly, this concern is not exclusive to SURF: semantic-entropy approaches generally depend on the reliability of semantic-equivalence estimation. SURF makes this dependence particularly important because singleton occupancy additionally informs the recalibration.

Thank you for the response. I appreciate the clarifications and the additional aggregate significance results. However, my main concerns remain only partially addressed.

  1. The full model improves over the fragility-only variant by only 0.0009 AUROC, and no correlation or paired comparison between full SURF and fragility-only SURF is provided.
  2. The paper still lacks clear guidance on when SURF should be preferred over simpler semantic-entropy baselines.
  3. The fragility signal is not directly validated.
  4. The end-to-end cost comparison remains incomplete.
  5. Robustness to clustering choices remains untested.

Overall, I appreciate the authors’ clarifications, but the response does not provide new empirical evidence for several central concerns. I am therefore inclined to keep my current rating.

AuthorsOpenReview ↗

Clarification of SURF’s Empirical Contributions and Responses to Remaining Concerns

We thank the reviewer for the follow-up and respectfully request reconsideration of the points addressed in our previous responses.

  • 1. AUROC and complementarity: The reviewer correctly identifies a +0.0009 difference between the mean AUROCs of full SURF and fragility-only SURF across 112 settings. However, this comparison does not fully reflect the contribution of the proposed method. SURF beats the SOTA sensitivity-aware Rényi SE (SE_RU, ICLR 2026) by introducing fragility, improving mean AUROC from 0.7732 to 0.7784 (+0.0052). Relative to Rényi SE, the improvement is +0.0089, with SURF outperforming it in 98/112 settings (Wilcoxon p = 2.2 × 10⁻¹³). These comparisons provide stronger evidence for the contribution of fragility-aware recalibration. We acknowledge that the additional benefit of combining UQ with fragility requires further statistical analysis.

  • 2. Practical applicability: As explained in W2, SURF is most relevant when multiple generations are already available and incomplete semantic coverage is a concern. Simpler approaches may be preferable in latency-constrained settings.

  • 3. Fragility validation: Q1 clarifies that fragility measures incomplete semantic coverage rather than factual incorrectness directly. The component ablation provides empirical evidence of its utility for hallucination discrimination.

  • 4. Computational cost: Q3 directs the reviewer to the supplementary computational analysis, which reports 6–10 ms per query for sensitivity estimation and 45–60 seconds for one-time Hamiltonian construction.

  • 5. Clustering robustness: Q4 explains that all evaluated methods use identical clustering conditions, enabling fair comparisons. We also acknowledge that alternative clustering methods warrant further investigation.

We respectfully encourage the reviewer to revisit these explanations and the existing experimental evidence when reconsidering their assessment.

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

Related papers

Discussion (0)

Sign in to comment.