acceptodds
Under review as a conference paper at ICLR 2027

Semantic Entropy Is Estimable Below the Recovery Threshold: Minimax Rates Under a Noisy Judge

Abstract

Semantic entropy scores a language model's uncertainty by the entropy of the distribution from which its sampled answers draw their meanings, and those meanings are never observed: an entailment judge is called on pairs and the answers are grouped by its verdicts. That grouping can fail when the judge is too weak for the available sample size. We model the judge as a channel that, given the meanings, declares each pair equivalent independently with probability when the two answers share a meaning and otherwise. For at most three meanings, a known channel with and a weakly separated judge, , we determine the minimax squared risk of the entropy: , with both constants depending only on . The rate is attained by an estimator that never groups the answers. It reads the collision probability off the agreement density and the triple-collision probability off a centred statistic in the squared degrees, projects the pair onto the set of valid moments, and solves a cubic. The matching lower bound is proved along a family that holds the agreement density fixed while the entropy changes, so every linear statistic of the verdicts has the same mean at its endpoints. With , the entropy is uniformly consistently estimable exactly for , and on the minimax risk is of order . For every , below the recovery threshold , no estimator of the meanings improves asymptotically on assigning each answer to the most likely one, even when given all other answers' meanings. On the entropy is therefore estimable while the grouping is unrecoverable.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
semantic entropy, uncertainty quantification, minimax rates, functional estimation, stochastic block model

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

Semantic entropy estimation typically involves (1) sampling nn times from a language model, (2) clustering the responses using a pairwise judge, and (3) estimating the entropy of the true distribution over clusters. In practice, this true distribution, its support size, etc. are not known, and several works have attempted to better estimate the functional of interest. Fewer works, on the other hand, are concerned with the pairwise judge (step 2). With this in mind, the authors study the minimax L2 risk of entropy estimation, where the judge/aggregator is noisy.

The paper itself is exceedingly dense, with somewhere in the vicinity of 4545 pages of proofs. There are several aspects the authors draw upon which are outside of my familiarity, and therefore it is intractable to thoroughly digest all these proofs in the time available, hence my confidence rating below. In my opinion, this kind of submission is better suited for a journal with longer review period. I am placing this comment here, because it does not necessarily pertain to the quality of the work presented and should not be treated as a weakness per se. That said, I do think it is quite crucial, as the paper is almost exclusively theoretical.

Critical strengths

S1: The topic is interesting and relevant. In fact, several things are undesirable about step 2 above (the pairwise judge) - for one, it is slow for large nn. For another, the judge may be bad or otherwise noisy (as the authors study), and if I remember correctly [1] observed that even humans did not fully agree on cluster assignments (and LLM judge agreed with human raters roughly as often as humans agreed with each other).

S2: Section 5.1, which identifies a region where consistent entropy estimation is impossible under a noisy judge, is interesting and the separation criterion seems sufficient to be a useful reference (i.e., the minimum separation for consistent estimation is not trivially small).


References

[1] Farquhar et al. "Detecting hallucinations in large language models using semantic entropy." Nature (2024).

Critical weaknesses

W1: Limited scope of applicability, for which semantic entropy calculated via (1) pairwise judge-based semantic aggregation, (2) where the true distribution over semantic clusters is drawn from a 3D probability simplex, and (3) where the judge's errors follow Eq. 1. I expect that this combination of constraints is quite rare, in practice.

  • Regarding (2): property estimation in general is especially challenging when the sample is smaller than the support, and this basically unaddressed by this work (and explicitly excluded in, e.g., Theorem 2, Corollary 3, etc.)
  • Regarding (3): Eq. 1 assigns the same clustering probability to all generations that truly do belong to the same cluster, which seems unrealistic. For instance, a judge based on an NLI model or good LLM should group together identical generations with probability at or near 1, but this would not be expected for more lexically distant, but still semantically equivalent, generations. Indeed, this is just a model, but it is incumbent on the authors to justify this model. Actually, it seems possible that this constraint could be more relevant outside the natural language generation setting (e.g., perhaps species identification has Bernoulli-esque noise).

W2: Abstruse writing throughout is a major obstacle, and this is independent of the comment I make in the Summary. For example, this statement from the abstract

"It reads the collision probability off the agreement density and the triple-collision probability off a centred statistic in the squared degrees, projects the pair onto the set of valid moments, and solves a cubic."

and this one from the introduction

"Its core is a path at constant agreement density, along which a Glauber identity and a blockwise Efron-Stein bound control the Fisher information of the graph despite the hidden meanings."

are extremely difficult to parse without inferring author intent or having already read the main body of the paper. In general, there are many unconventional, undefined, or otherwise confusing aspects throughout (e.g., Sections 4.2 and 5.1 allude to an unexplicated "Glauber identity" and Laplacian-K is only defined in Appendix G.3) that should be clarified in revision.

W3: I am cautious of some of the mathematical content, apropos of the caveat included in the Summary section. To start, it is not obvious to me that P^3\hat{P}_3 is an unbiased estimator of P3P_3 (L188). I suspect that P^3\hat{P}_3 may be a negatively biased, but consistent, estimator of P3P_3, though I could be wrong. Can the authors elaborate on this? For simplicity, let's take a noiseless judge (i.e., q=0q=0, Δ=1\Delta=1) and θ=(0.5,0.5,0)\theta = (0.5, 0.5, 0). Also, let YY denote the set of integers in [2,n][2, n], which I am struggling to typeset here. Then P3=0.25P_3 = 0.25, and, if I am correct, P^3∈(i−1)(i−2)(n−1)(n−2):i∈Y\hat{P}_3 \in \\{ \frac{(i-1)(i-2)}{(n-1)(n-2)}: i \in Y \\}. For my curiosity, perhaps the authors can run a large number of Monte Carlo trials illustrating this scenario with some small values of n>3n>3, e.g., n∈[4,10]n \in [4, 10].

I have similar feelings elsewhere. I am not going to be exhaustive here, as the guidance suggests these explanations be brief, but I feel a point-by-point review of the proofs is warranted, and W2 makes this especially challenging.

Questions for authors

Please see the Weaknesses, where I have provided many details.

  1. Can the authors please clarify which presented items are original and which are drawn from prior work? For instance, the related work says "The two identities we use appear in Ambroise & Matias (2012)" - which two identities (does this refer to Eq. 4)?

  2. Can the authors please review their notation for conceptual overlap? E.g., Δ\Delta (including with subscript) refers to both the probability simplex and the judge's separation).

  3. Is Figure 2 correctly labeled?

2 comments
AuthorsOpenReview ↗

Response to Reviewer DXdR (1/2)

W3. With the true channel, P^3\widehat P_3 is exactly unbiased for every n≥3n\ge3. Write di−q(n−1)=∑j≠i(Aij−q)d_i-q(n-1)=\sum_{j\ne i}(A_{ij}-q) and expand the square. As Aij∈{0,1}A_{ij}\in\lbrace0,1\rbrace, the diagonal terms sum to di(1−q)2+(n−1−di)q2d_i(1-q)^2+(n-1-d_i)q^2, exactly what Equation (4) subtracts. So P^3\widehat P_3 is the average of (Aij−q)(Aik−q)/Δ2(A_{ij}-q)(A_{ik}-q)/\Delta^2 over ordered triples of distinct answers. Given the meanings, the two verdicts are independent with E[Aij−q∣Z]=Δ1{Zi=Zj}\mathbb E[A_{ij}-q\mid Z]=\Delta\mathbf 1\lbrace Z_i=Z_j\rbrace, so each term has mean P(Zi=Zj=Zk)=P3\mathbb P(Z_i=Z_j=Z_k)=P_3 (Lemma B.1).

Thank you for the checkable example (q=0q=0, Δ=1\Delta=1, θ=(0.5,0.5,0)\theta=(0.5,0.5,0)). There di=nZi−1d_i=n_{Z_i}-1, with ncn_c the number of answers of meaning cc, so P^3=∑cnc(nc−1)(nc−2)/(n(n−1)(n−2))\widehat P_3=\sum_cn_c(n_c-1)(n_c-2)/(n(n-1)(n-2)). Since n1,n2∼Bin(n,12)n_1,n_2\sim\mathrm{Bin}(n,\tfrac12) have E[nc(nc−1)(nc−2)]=n(n−1)(n−2)/8\mathbb E[n_c(n_c-1)(n_c-2)]=n(n-1)(n-2)/8, EP^3=14=P3\mathbb E\widehat P_3=\tfrac14=P_3. Your (i−1)(i−2)/((n−1)(n−2))(i-1)(i-2)/((n-1)(n-2)) is the contribution of one answer whose meaning is shared by ii answers. P^3\widehat P_3 averages it over all nn answers: at n=4n=4 it takes the values 1,14,01,\tfrac14,0 with probabilities 216,816,616\tfrac2{16},\tfrac8{16},\tfrac6{16}, with mean 14\tfrac14.

Evaluating Equation (4) literally in your example, 10610^6 trials per n=4,…,10n=4,\dots,10 give means 0.24971 to 0.25049, all within two standard errors of 14\tfrac14. Exact enumeration gives 14\tfrac14 at each nn, also at n=4,5,6n=4,5,6 with q=0.3q=0.3, Δ=0.2\Delta=0.2. The table is in our comment (2/2) and will go into the appendix.

Proofs. The qualitative separation uses only Theorem 2 (Appendix B) and Proposition 8 (Appendix E.2, about one page). Only the matching lower bound and the per-answer statement of Proposition 7 need Appendix D. Our comment (2/2) gives a reading route.

Q1. Yes: Equation (4) uses E[Aij]=q+ΔP2\mathbb E[A_{ij}]=q+\Delta P_2 and E[(Aij−q)(Aik−q)]=Δ2P3\mathbb E[(A_{ij}-q)(A_{ik}-q)]=\Delta^2P_3. Ambroise and Matias (2012) give these identities and estimate unordered proportions without recovering labels. Liu (2026) bounds the entropy error via clustering, with no lower bound, and projection before inversion is prior (Wu and Yang, 2020). We prove matching entropy-risk bounds as the judge weakens, the exact consistency threshold, and entropy estimation where recovery is impossible (Theorems 2, 4 and 5, Propositions 6 to 8). The entropy bound includes the simplex boundary. We will state this split in the related work.

Q2, Q3. We will write the simplex as S3\mathcal S_3, reserving Δ\Delta for the judge, and give the class counts of Appendix D their own symbols. Figure 2 is correctly labelled but under-explained: its axes are e2=(1−P2)/2e_2=(1-P_2)/2 and e3=(1−3P2+2P3)/6e_3=(1-3P_2+2P_3)/6, and the triples printed in the plane are the distributions θ\theta at those points, not coordinates. We will say so in the caption and label the points θ=(1,0,0)\theta=(1,0,0) etc.

W2. We agree and will replace both quoted sentences. The estimator infers the probabilities that two or three answers share a meaning, then computes the entropy from them. The lower bound compares distributions with different entropies but equal expected edge counts, bounding the Fisher information along a path between them. We will state the Glauber identity (Lemma D.8), the conditional law of one meaning given the graph and all other meanings, on first use in Section 4.2. Figure 1's caption will define Laplacian-KK (Liu, 2026), which groups answers by the eigenvectors of the KK smallest eigenvalues of D−AD-A, DD being the diagonal degree matrix and KK the true number of meanings.

W1. Scope. Equation (1), as in Liu (2026), isolates the cost of not observing the meanings. With observed labels the entropy on a fixed alphabet has minimax rate 1/n1/n, and seeing the labels only through a weak judge costs a factor n2γ−1/2n^{2\gamma-1/2} for Δ=n−γ\Delta=n^{-\gamma}, 1/2<γ<3/41/2<\gamma<3/4 (Section 2). We have no matched rate when the support can exceed nn, and we have not established all three constraints jointly for real judges. Under the stated hypotheses, the lower bounds extend to any support bound K≥3K\ge3 and to larger classes containing Equation (1) with the same graph data and entropy target. Inclusion alone does not extend the upper-bound guarantee.

We tested four entailment models on 80 SciQ questions, each with three candidate words in 12 templates. Meaning is candidate identity, ignoring confidence wording. All four judge identical strings equivalent. With all 12 templates, mean bidirectional agreement on distinct wordings of the same candidate ranges from 0.639 to 0.943. So the entailment models show the contrast you describe, for wording around a fixed word rather than distant paraphrases. Across 16 configurations (4 models, 2 rules, 2 template sets), calibration on 40 other questions gives the pooled diagnostic nΔ^2≥4.58n\hat\Delta^2\ge4.58 at n=10n=10, far outside the matching lower-bound regime (nΔ2≤c1<0.01n\Delta^2\le c_1<0.01). These measurements do not validate the theorems and will go into the appendix.

AuthorsOpenReview ↗

Response to Reviewer DXdR (2/2)

Details behind our comment (1/2).

The check of P^3\widehat P_3 in your example. Equation (4) is evaluated literally from the degrees. "Exact" sums over every noiseless labelling. For the noisy channel it groups labellings by class size with their binomial weights and enumerates all graphs for each size. "MC" is the mean over independent trials, with its standard error in parentheses.

nnnoiseless, exactnoiseless, MC (10610^6 trials)q=0.3q=0.3, Δ=0.2\Delta=0.2, exactq=0.3q=0.3, Δ=0.2\Delta=0.2, MC (4×1054\times10^5 trials)
41/40.25049 (0.00031)1/40.2465 (0.0032)
51/40.25046 (0.00024)1/40.2526 (0.0021)
61/40.24971 (0.00019)1/40.2517 (0.0016)
71/40.24989 (0.00016)not enumerated0.2504 (0.0013)
81/40.24988 (0.00014)not enumerated0.2501 (0.0011)
91/40.25017 (0.00013)not enumerated0.2498 (0.0009)
101/40.25003 (0.00011)not enumerated0.2508 (0.0008)

The two quoted sentences in plain terms.

The abstract sentence. The estimator uses two summary statistics of the verdicts. Subtracting qq from the fraction of pairs judged equivalent and dividing by Δ\Delta estimates P2P_2, the probability that two answers share a meaning. A corrected average of the squared degrees, divided by Δ2\Delta^2, estimates P3P_3, the probability that three answers share a meaning. The degree of an answer is the number of answers it is judged equivalent to. Noise can put (P^2,P^3)(\widehat P_2,\widehat P_3) outside the moment pairs of distributions on at most three meanings, and we move it to the nearest such pair. With at most three meanings the proportions are then the roots of a cubic, and their entropy is the estimate.

The introduction sentence. We compare two distributions with different entropies and the same expected number of agreeing pairs, joined by a path of distributions that keeps that expectation fixed. The Fisher information of the graph along the path bounds the total variation distance between the two graph laws, which limits how reliably any estimator can tell them apart. What we called the Glauber identity (Lemma D.8) is Bayes' rule for one answer: given the graph and all other meanings, multiply each meaning's prior weight θc\theta_c by the likelihood of that answer's verdicts if it had meaning cc, and normalize. With it, Lemma D.10 expresses the graph score (the derivative of its log-likelihood along the path) as the conditional mean, given the graph, of an explicit mean-zero function of the graph and meanings. This function's second moment bounds the Fisher information. Efron–Stein bounds this second moment by half the sum of the expected squared changes when one edge's randomness or one answer's meaning is resampled at a time.

A reading route for the proofs.

  • The estimator's guarantee (Theorem 2). Appendix B, pages 12 to 19, self-contained.
  • Whole-grouping separation. Under Assumption 1, Δ=n−γ\Delta=n^{-\gamma}, 1/2<γ<3/41/2<\gamma<3/4, Theorem 2 gives uniform entropy consistency. Proposition 8 rules out recovery up to relabelling for fixed non-point-mass θ\theta (Appendix E.2, about one page).
  • The per-answer statement with an oracle (Proposition 7). This also uses the posterior identity and several auxiliary bounds from Appendix D.
  • No uniformly consistent estimator unless n3Δ4→∞n^3\Delta^4\to\infty (Proposition 6). Appendix C, pages 19 to 27.
  • The matching lower bound (Theorem 4). All of Appendix D, pages 27 to 45. Figure 4 is its roadmap.

The appendix also reports numerical checks of selected pieces. Over 216 simulated combinations, the means of P^2−P2\widehat P_2-P_2 and of P^3−P3\widehat P_3-P_3 are within 2.3 standard errors of their predicted values (Appendix G.4). These checks test the implementation and selected identities, not the uniform bounds, and they do not replace reading the lower-bound proof. We would be glad to go through any step you would like checked.

Measured bias under the real judges. Two equivalence rules (bidirectional entailment, or the default of the released code of Farquhar et al., 2024) and two template sets (all 12, or the 7 without first-person or modal markers) give 16 configurations for the four entailment models. With the channel calibrated on 40 other questions, the exact biases of P^3\widehat P_3 under fixed verdicts and uniform template sampling, averaged over the six assignments of each of the four distributions of Table 1 to the candidates, range from −0.004-0.004 to +0.029+0.029 after averaging over questions and from −0.23-0.23 to +0.26+0.26 for individual questions. A language-model judge, Qwen2.5-14B-Instruct with the entailment prompt of that released code, behaves differently. Across its four configurations the test-set mean agreement rates are 0.932 to 0.966 for the correct candidate and 0.075 to 0.127 for the same distractor. This is class-dependent behaviour, which the Limitations name as not covered.

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

This paper is on semantic entropy, which characterizes LLM uncertainty by clustering the answers generated with respect to their shared "meaning", and then estimating the entropy with respect to the different meaning classes. The grouping of the answers with respect to their meanings is done by a "judge LLM", which takes pairs of answers and decides whether they entail the same meaning. The paper aims to model the "judge" in a semantic estimator as a noisy channel, and derive minimax squared risk of the entropy.

Critical strengths

Uncertainty estimation is an important problem, and understanding the theoretical limits of semantic estimators would be a nice contribution.

Critical weaknesses
  1. For a reviewer who has background in information theory but not necessarily on semantic entropy, I found the paper presentation, technical arguments not clear and very hard to follow, which makes the contributions very hard to assess.

  2. The paper does not motivate the technical problem clearly. There is a lot of technical jargon, starting from the introduction and problem setup, but the paper does not give much intuition behind these design choices. Starting from the abstract, for instance, the paper refers to a weakly separated judge "n\Delta^2 \leq c_1", however, does not discuss what n, \Delta or c_1 is. These are only defined at the end of the intro. The abstract should be self-contained. This general trend is followed throughout the paper, please see the details below.

  3. The introduction should motivate the problem and contributions better. I had a very hard time to following the contributions. In the third paragraph (in introduction), for instance, what do you mean "the entropy needs only the proportion of meanings"? And P2, P3 appear out of nowhere, P_k is only defined after equation 2 later in Section 3. Also, what is a "triple-collision probability"? Similarly the intro refers to a verdict graph reducing to a stochastic block model, and one of the contributions listed in the intro is "The rate is attained by a projected two-moment estimator that reads the graph once and solves a cubic, with no tuning parameter and no spectral step; its upper bound holds with explicit constants at every n and every channel.". But it is not discussed what is a verdict graph, or what is referred to as a "spectral step". These have to be later guessed from later sections.

  4. Starting from Section 3 (titled "Setup"), the first paragraph (Model): We are told Z1, \ldots, Zn are independent and each have a probability of having a realization c with probability \theta_c. We are not told what "c" is, or what "Z_i" is. Is it representing the meaning, or the answers? What is the judge trying to assess, whether Z_i and Z_j have the same meaning? I had to infer these by looking up from Kuhn's paper, so I am guessing "c" represents a semantic equivalence class sharing a "meaning". However, the paper should be self-contained and should clearly state the key variables during the problem setup and methodology. More importantly, it is not clear how they relate to the estimators formally, as I describe in the next item.

  5. Continuing on the problem setup, parts that I could not decipher: What is the formal definition of \widehat{H}? How is equation (1) related to equation (2)? How is A_{ij} related to \theta_c and accordingly H(\theta)? I am very confused how the judge verdicts are formally incorporated in the estimator.

  6. Why is the channel transition probability "q" known? I find this not very convincing. What is the practical intuition behind this? How do we know q+ \Delta is a valid Bernoulli parameter? And it is also mentioned in the contributions section (end of the Introductions) that "The channel can also be calibrated from labelled pairs at no cost in rate". How is this calibration done exactly and why does it cost no rate?

  7. I could not tell what the caption of Figure 1 is trying to convey. What do you mean "Where the semantic entropy is estimable and where the meaning partition is not.". There is too much technical jargon but no intuition.

  8. In Assumption 1, it is mentioned that labels are iid from \theta. What is the label referred to here?

  9. And after the unclear problem setup the results are listed one after the other, followed by the experiments. I went through each part, however, since at the current stage the problem formulation and methodology is not clear, I cannot confidently validate the technical contributions.

Questions for authors

Please see my comments raised in the weaknesses. The paper should clearly/formally define the technical problem and methodology, and provide clear intuition behind the design/modeling choices. Currently it appears more like a technical report with a lot of technical jargon and incomplete and/or poorly defined technical arguments. I feel that a lot of re-work should be done to make it publishable.

2 comments
AuthorsOpenReview ↗

Response to Reviewer PiEm (1/2)

We agree that the setup needed clearer definitions.

W1–W3. The question. If two meanings have probabilities 0.8 and 0.2, two independent answers share a meaning with probability 0.82+0.22=0.680.8^2+0.2^2=0.68. The entropy also depends only on these proportions, not on which answer has which meaning. Can noisy comparisons reveal enough about the proportions when individual answers cannot be reliably grouped?

W4, W5, W8. The model. For one question, the n≥3n\ge3 answers have hidden meanings ZiZ_i, independent draws from θ\theta. Here cc indexes a meaning, θc=Pr⁡(Zi=c)\theta_c=\Pr(Z_i=c) for c=1,2,3c=1,2,3, and zero masses are allowed. The ZiZ_i are the labels of Assumption 1. We observe Aij=1A_{ij}=1 if the judge calls answers i,ji,j equivalent and 00 otherwise. Given the meanings, verdicts are independent. Agreement has probability q+Δq+\Delta for a shared meaning and qq otherwise, with 0≤q<q+Δ≤10\le q<q+\Delta\le1 (Theorem 2), so q+Δq+\Delta is a valid Bernoulli parameter. The difference Δ\Delta is the judge's separation in Equation (1). The verdict graph has one vertex per answer and one edge per agreement. Its edge probabilities depend only on the hidden groups of the endpoints, which makes it a stochastic block model.

Our target is the population entropy H(θ)=−∑cθclog⁡θcH(\theta)=-\sum_c\theta_c\log\theta_c. For the graph AA drawn under Equation (1), Equation (2) is the minimax risk inf⁡gsup⁡θEθ[(g(A)−H(θ))2]\inf_g\sup_\theta\mathbb E_\theta[(g(A)-H(\theta))^2] over estimators gg given the channel.

From verdicts to entropy.

  1. The fraction of agreeing pairs has mean q+ΔP2q+\Delta P_2. Subtract qq and divide by Δ\Delta to estimate the collision probability P2=∑cθc2P_2=\sum_c\theta_c^2, the probability that two draws share a meaning.
  2. For distinct i,j,ki,j,k, (Aij−q)(Aik−q)/Δ2(A_{ij}-q)(A_{ik}-q)/\Delta^2 has mean P3=∑cθc3P_3=\sum_c\theta_c^3, the triple-collision probability that three draws share a meaning. Equation (4) averages these products using each answer's number of agreements.
  3. With at most three meanings these two moments determine the unordered probabilities. Noise can give a pair of moments that no distribution on at most three meanings has. Algorithm 1 uses a closest valid pair in Euclidean distance, solves a cubic for the corresponding probabilities θ~\tilde\theta, and returns H^=H(θ~)\widehat H=H(\tilde\theta).

W2, W7, W9. Figure 1. A weak judge's signal in one answer's agreement count is small compared with its noise. We write nΔ2≤c1n\Delta^2\le c_1, where the constant depends only on how far qq stays from 00 and 11. Panel (a) illustrates Δ=n−2/3\Delta=n^{-2/3} at n=500n=500 and 20002000, where nΔ2n\Delta^2 is 0.13 and 0.08, above c1c_1. Algorithm 1 follows the changing entropy. The two methods that first group the answers using eigenvectors of the centred verdict matrix or of D−AD-A, with DD the diagonal matrix of agreement counts (the spectral step), each give a nearly constant mean estimate.

Panel (b) explains the exponent for a known channel with qq bounded away from 00 and 11. At Δ=n−2/3\Delta=n^{-2/3}, nΔ2=n−1/3→0n\Delta^2=n^{-1/3}\to0. For fixed θ\theta with at least two positive masses, the smallest expected fraction misgrouped after relabelling tends to 1−max⁡cθc1-\max_c\theta_c, the same limit as putting all answers in one group. Yet n3Δ4=n1/3→∞n^3\Delta^4=n^{1/3}\to\infty, enough for uniform entropy consistency (Proposition 8 and Theorem 2). Section 4.2 sketches the matching lower bound: it compares two distributions with equal expected edge counts and different entropies and controls the distance between their full graph laws.

W6. The channel. Algorithm 1 takes (q,Δ)(q,\Delta) as input. In practice they are measured on labelled pairs, and "no cost in rate" means the same risk order in (n,Δ)(n,\Delta), at the price of these pairs and a larger constant. Estimate qq by the agreement fraction on mm different-meaning pairs and p=q+Δp=q+\Delta on mm same-meaning pairs, and set Δ^=p^−q^\hat\Delta=\hat p-\hat q. For nΔ2≤c1n\Delta^2\le c_1 and n3Δ4≥4n^3\Delta^4\ge4, Corollary 10 keeps the upper-bound order for fixed calibration values satisfying ∣q^−q∣+∣Δ^−Δ∣≤1/(n3/2Δ)|\hat q-q|+|\hat\Delta-\Delta|\le1/(n^{3/2}\Delta). With mm labelled pairs per kind, the expected sum of these errors is at most 3/(2m)3/(2\sqrt m), and Corollary F.4 gives the same upper-bound order after averaging over calibration whenever nΔ2≤c1n\Delta^2\le c_1 and m≥94n3Δ2m\ge\tfrac94 n^3\Delta^2. The estimator returns 00 if Δ^=0\hat\Delta=0. That the pairs share the target channel and are judged independently of one another and of AA is an assumption. With qq unknown in [q0,1−q0][q_0,1-q_0], 0<q0<1/20<q_0<1/2, verdicts alone cannot give uniform consistency in the weak regime even with Δ\Delta known (Proposition F.5), so some information on qq is necessary.

We will define these terms where the introduction first uses them, open Section 3 with this model, the parameter range and the estimator, explain the calibration before Corollary 10, and write "meanings" for "labels" throughout. Our comment (2/2) gives the revised abstract and Figure 1 caption.

AuthorsOpenReview ↗

Response to Reviewer PiEm (2/2)

We will use the following abstract and Figure 1 caption, and Section 3 will open with the model and the three steps of our comment (1/2). The theorem statements do not change. Under the assumptions and regime below, entropy estimation where recovery is impossible follows from Theorem 2 (Appendix B) and Proposition 8 (Appendix E.2, about one page). The matching lower bound and the per-answer statement of Proposition 7 also need Appendix D.

Revised abstract.

Semantic entropy is the Shannon entropy of the distribution over the meanings of a language model's answers to one question. We ask whether this entropy can be estimated when a noisy judge makes individual answers hard to group. For one question we sample nn answers whose meanings are independent draws from a distribution θ\theta. Given the meanings, the judge's verdicts on pairs are independent. The probability of agreement is q+Δq+\Delta for two answers with the same meaning and qq otherwise. The gap Δ\Delta measures the judge's ability to separate the two cases.

With at most three meanings and a known channel, we estimate the entropy by counting agreements and pairs of agreements that share an answer, without grouping the answers. When qq stays away from 00 and 11 and nΔ2n\Delta^2 is sufficiently small, the worst-case mean squared error of this estimator is of the minimax order min⁡{1,(n3Δ4)−1/2}\min\lbrace1,(n^3\Delta^4)^{-1/2}\rbrace. Uniform consistency holds exactly when n3Δ4→∞n^3\Delta^4\to\infty. For Δ=n−γ\Delta=n^{-\gamma} with 1/2<γ<3/41/2<\gamma<3/4, the error therefore vanishes. There, at any fixed θ\theta with at least two positive probabilities, the smallest expected fraction of answers misgrouped after relabelling has the same limit as for the rule that puts every answer in one group. Even revealing all other meanings leaves the same limiting error for a single answer as always guessing the most probable meaning. The entropy is thus estimable where the grouping is not recoverable.

Revised caption of Figure 1.

(a) For four distributions, the horizontal axis gives the true entropy and the vertical axis the mean estimate in the six settings of Table 1, at Δ=n−2/3\Delta=n^{-2/3}. Algorithm 1 follows the changing entropy. Spectral clustering, using the three leading eigenvectors of the centred verdict matrix, stays near log⁡3\log 3. Laplacian-KK clustering (Liu, 2026), using the KK lowest-eigenvalue directions of D−AD-A, stays near 0. Here AA is the verdict matrix, DD is the diagonal matrix of agreement counts, and KK is the true number of meanings.

(b) For at most three meanings, a known channel and qq bounded away from 0 and 1, let Δ=n−γ\Delta=n^{-\gamma}. The exponent rr measures how fast the minimax squared error falls, as n−rn^{-r}. For γ>1/2\gamma>1/2 and any fixed θ\theta with at least two positive probabilities, the smallest expected fraction of answers misgrouped after relabelling has the same limit as for putting all answers in one group. On 1/2<γ<3/41/2<\gamma<3/4, entropy can still be estimated with r=(3−4γ)/2>0r=(3-4\gamma)/2>0. At γ≥3/4\gamma\ge3/4, even uniform entropy consistency fails. The star is the exponent 2/32/3 used in (a). Those finite sample sizes lie outside the matching lower bound's small-constant regime.

Reviewer_RN3VOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper shows very interesting theoretical results that even when the clusters of semantic meanings is not recoverable due to a noisy judge, the semantic entropy is estimable. The algorithm is simple; it requires no hyperparameters. However, it does require the judeg/channel properties such as q and Δ\Delta.

Critical strengths

Strengths:

  1. It challenges the assumption that clustering is needed to estimate semantic entropy.
  2. Gives non-asymptotic upper bound. The lower bound’s constant is not evaluated which the authors acknowledge as a limitation.
  3. The algorithm is simple: just count the edges, squared degrees and solve a cubic.
Critical weaknesses

Weaknesses:

  1. The authors acknowledge that the study is limited to at most K=3 meanings, which limits the work’s applicability. The authors provide no real-world justification for where this assumption would be applicable. In regimes where LLMs hallucinate, is K≤3K\leq3 sufficient?
  2. Most experiments are on synthetic experiments that assume the proposed channel model. Having real-world experiments using LLMs or real NLI judges would be very insightful.
Questions for authors

See above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer RN3V (1/2)

W1. Is K≤3K\le3 enough under hallucination? Not in general. Three is the first case where agreement density does not determine entropy, and the last where two moments do (Section 5.2), so the estimator does not extend beyond it. With 10 OPT-30B answers per question, Kuhn et al. (2023, Table 2) report on average 1.27 and 1.77 judge-formed clusters for correctly and incorrectly answered CoQA questions, and 1.89 and 3.89 for TriviaQA. These sample averages do not bound the number of meanings of the underlying distribution, but the last exceeds three. We will state this limitation and these numbers beside the assumption. For any support bound K≥3K\ge3, padding a three-meaning distribution with zeros preserves entropy and the verdict law, so the lower bound of Corollary 3 and the necessity of n3Δ4→∞n^3\Delta^4\to\infty for uniform consistency (Proposition 6) transfer under their hypotheses. We will add this as a remark.

W2. Real judges. Thank you. Our controlled check of Algorithm 1 finds heterogeneous verdict rates and relative finite-sample performance that depends on the distribution, templates and baseline. It does not validate the minimax rate or recovery threshold. We will add the study with all tables to the appendix and point to it from Section 6.

Setup. SciQ (Welbl et al., 2017): 80 test and 40 disjoint calibration questions, each with three candidate words (one correct, two distractors) in 12 fixed templates. Meaning is candidate identity, ignoring confidence wording. Fixed verdicts come from four entailment models and from Qwen2.5-14B-Instruct with the entailment prompt of Farquhar et al. (2024). Two rules (bidirectional entailment as in Kuhn et al., or the default of Farquhar et al.'s code) crossed with two template sets (all 12, or the 7 without first-person or modal markers) give 16 entailment-model and 4 Qwen configurations. We calibrate (q^,Δ^)(\hat q,\hat\Delta) on the 40 questions. For each Table 1 distribution and test question, 100 repetitions uniformly permute θ\theta across candidates and sample templates uniformly with replacement, at n∈{10,20,50,100,200}n\in\lbrace10,20,50,100,200\rbrace. Risk and bias averages need not describe performance for a fixed assignment.

  1. Estimation error (exploratory). Estimated mean squared errors (nats squared) average over the same 80 questions and random candidate assignments at fixed calibration. Under bidirectional entailment with all templates, Algorithm 1 has lower error than our implementation of the greedy clustering of Farquhar et al. with plug-in entropy in all 80 cells (4 entailment judges, 4 distributions, 5 sample sizes), by 0.019 to 0.383. Clipping the greedy estimate at log⁡3\log3 enforces the same entropy cap: Algorithm 1 is then lower in all 40 near-vertex and boundary cells but higher in 34 of the 40 interior and uniform cells, by at most 0.023. With the default rule and 7 templates, where the same-candidate rate is at least 0.996, it and unclipped greedy differ by at most 0.008 for n≥20n\ge20. At n=200n=200, Laplacian clustering given the true number of meanings is lower in 36 of 64 cells (16 configurations, 4 distributions).
  2. Outside the weakly separated regime. Across the 16 configurations, the pooled diagnostic nΔ^2n\hat\Delta^2 is at least 4.58 at n=10n=10, while this regime requires nΔ2≤c1n\Delta^2\le c_1 with c1<0.01c_1<0.01.
  3. The channel is not homogeneous. All four entailment judges declare identical strings equivalent. Under the bidirectional rule with all templates, two different templates naming the same candidate are declared equivalent at mean rates of 0.639 to 0.943 across the four judges. Over the 16 configurations, the three same-candidate rates of one question differ by up to 0.41. Averaged over assignments of θ\theta to candidates, the exact bias of P^3\widehat P_3 with the calibrated channel is −0.004-0.004 to +0.029+0.029 on average over test questions (16 configurations, four distributions) and −0.23-0.23 to +0.26+0.26 for individual questions.
  4. The language-model judge is outside our model. Its agreement depends on whether the candidate is correct: across its four configurations the mean rates are 0.932 to 0.966 for two answers naming the correct candidate and 0.075 to 0.127 for two naming the same distractor. The Limitations name such class-dependent rates as not covered, and the mean bias of P^3\widehat P_3 is then +0.148+0.148 to +1.267+1.267.

Limits. The task is semi-synthetic: one dataset, templated answers and a prescribed θ\theta, not a language model's own answers. Algorithm 1 is given the calibrated channel and the three-meaning bound, and unclipped greedy uses neither.

The channel and the lower-bound constant. Calibration from labelled pairs assumes they share the target channel (Corollary F.4). With qq unknown in [q0,1−q0][q_0,1-q_0], 0<q0<1/20<q_0<1/2, uniform consistency from verdicts alone over (q,θ)(q,\theta) fails in the weakly separated regime (Proposition F.5). The constant of Theorem 4 is finite and not evaluated, as the Limitations state.

AuthorsOpenReview ↗

Response to Reviewer RN3V (2/2)

This comment gives the details of the study summarized in our comment (1/2).

Judges and rules. The entailment models are DeBERTa-large-MNLI, DeBERTa-v2-xlarge-MNLI, a DeBERTa-v3-large NLI model and RoBERTa-large-MNLI. Qwen2.5-14B-Instruct is run with the entailment prompt and parsing rule of the released code of Farquhar et al. (2024). Labels become verdicts by one of two rules: entailment in both directions (Kuhn et al., 2023), or the default of that released code, no contradiction in either direction and not neutral in both. The templates range from "It is c." to "I believe the answer is c.", and the 7 plain ones have no first-person or modal marker.

The measured channel. Over the 16 configurations of the entailment judges, two answers naming different candidates are judged equivalent at calibrated rate q^\hat q between 0.0015 and 0.0143, and two naming the same candidate at rate p^\hat p between 0.683 and 1.000. Under bidirectional entailment with all templates, two different templates naming the same candidate are declared equivalent at rates of approximately 0.764, 0.819, 0.639 and 0.943 for the four judges, and at 0.953 to 1.000 with the plain templates. For individual questions a judge can conflate two different candidates: RoBERTa-large under the second rule judges every answer naming cratons equivalent to every answer naming craters.

The two statistics. Conditional on the fixed verdict tables and on uniform sampling of templates, and averaged over the six assignments of θ\theta to the candidates and over the test questions, the bias of P^2\widehat P_2 under the calibrated channel is between −0.020-0.020 and +0.024+0.024 and that of P^3\widehat P_3 between −0.004-0.004 and +0.029+0.029, over the 16 configurations and the four distributions. Signed averages can hide cancellation: for individual questions the bias of P^3\widehat P_3 ranges from −0.23-0.23 to +0.26+0.26. The covariance contribution of verdicts that share an answer, ζ/Δ^2\zeta/\hat\Delta^2 in the notation of Proposition F.6, averages between 0.00008 and 0.042, and is at most 0.0041 with the plain templates. Proposition F.6 assumes homogeneous means, which these judges do not have, so this is a measured decomposition and not an application of the proposition.

Errors. Under bidirectional entailment with all templates, the advantage of Algorithm 1 over unclipped greedy clustering exceeds six standard errors of the paired difference in each of the 80 cells. At n=200n=200 the estimated mean squared error of Algorithm 1 for the entailment judges is between 0.00009 and 0.025 with the calibrated channel, against at most 0.004 for the oracle that sees the meanings and at most 0.015 with the question's own channel, a diagnostic that uses the ground truth of the test question. These comparisons show residual finite-sample error and sensitivity to the channel. They do not isolate calibration transfer as the only cause.

The language-model judge. Mean agreement on identical strings is 0.415 to 0.444. Across four configurations, four distributions and five sample sizes, mean squared error is 0.037 to 0.363 for Algorithm 1, 0.58 to 13.4 for greedy clustering (at most 0.367 when clipped) and 0.011 to 0.709 for Laplacian clustering. For this model and prompt, these errors measure performance outside the homogeneous-channel assumptions.

Limits. The task is controlled and semi-synthetic, on one dataset: templated short answers, three candidates and a prescribed θ\theta, not a language model's own answers. Judges are deterministic, and samples draw uniformly within each candidate's templates, with replacement, from 36 strings in total, or 21 with the plain templates. Each repetition randomly permutes θ\theta across the candidates. Candidate triples were screened using WordNet rules and two language models, and we have checked the retained triples. The methods do not receive the same information: Algorithm 1 has the calibrated channel and the bound of three meanings, the clipped greedy estimate that bound only, the Laplacian clustering the true number of meanings, greedy clustering nothing. Greedy clustering is our reimplementation of the released algorithm, which judges repeated strings like any other pair, whereas the code of Kuhn et al. first merges identical strings. Standard errors are across test questions and exclude the uncertainty of the calibration set, and the configurations reuse the same questions. The two rules, the plain subset and the prompt were fixed after pilots on a few questions, which were then excluded. After a first complete analysis, greedy clustering was changed to follow the released algorithm exactly, and the clipped comparison and the per-question channel diagnostics were added, so these comparisons are exploratory. All 20 configurations will be reported in the appendix.

Reference: J. Welbl, N. F. Liu and M. Gardner. Crowdsourcing multiple choice science questions. Workshop on Noisy User-generated Text, 2017.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.