Pith. sign in

REVIEW 5 major objections 6 minor 9 references

Ensemble based approach to quantifying uncertainty of LLM based classifications

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper argues that under greedy decoding, variance in an LLM's classification is governed by conceptual certainty and lexical input variance, and that paraphrase-vote counts can estimate whether a prediction is correct.

desk verdict Plausible idea, but the evidence is in-sample: the paper never tests whether its ensemble-accuracy score predicts correctness on new questions. read the letter →

arxiv 2502.08631 v2 pith:K7YHUWIT submitted 2025-02-12 cs.AI

classification cs.AI
keywords LLMuncertaintyparaphraseensemblemajority-votecertaintyfine-tuningclassificationconfidencegreedydecodingconceptualempiricalCDF
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that when an LLM answers greedily, the variation in its predicted class across differently worded versions of the same question is controlled by two things: how solid the model's parametric knowledge is for that concept, and how much the input wording varies. Fine-tuning, the paper argues, raises the first quantity, so a fine-tuned model's classifications become more stable under paraphrasing. The paper turns this idea into a working uncertainty estimate by generating 15 paraphrases per question, taking the majority-vote frequency as the prediction's certainty, and showing that this score separates correct from incorrect predictions on 179 labeled questions (Kolmogorov-Smirnov statistic 0.809). The practical payoff is a black-box way to attach a probability of correctness to a new classification, using only rephrased inputs and the model's own outputs, which matters in settings where a wrong answer cannot be silently accepted.

What carries the argument

The machinery is the paraphrase ensemble. For each latent intent $Q_i$ (the underlying meaning of a natural-language question), the paper generates 15 surface variants $Q_{ij}$ with an LLM, runs greedy decoding on each, and records the predicted endpoint or parameter list. The identity doing the work is Equation (2), which says that variance in the predicted class is a function of conceptual certainty and input variance; the practical score derived from it is ensemble accuracy, the normalized frequency of the majority class. That score is then compared against two empirical cumulative distribution functions (ECDFs), one built from correct predictions and one from incorrect predictions, and the tail probabilities of those ECDFs give the correctness estimate. The Kolmogorov-Smirnov test is the evidence that the two distributions are not the same.

What would settle it

Collect a larger labeled set with many more incorrect predictions, recompute the KS statistic between the ensemble-certainty distributions for correct and incorrect predictions, and check whether the separation seen here ($0.809$) persists; if the small incorrect sample is driving the separation, the method's correctness signal disappears.

Watch

Extended reading notes

Core claim

The central claim is Equation (2): $\mathrm{var}(\mathrm{LLM\ classification}) = F(\mathrm{conceptual\ certainty}, \mathrm{var}(\mathrm{LLM\ input}))$. In words, once decoding is greedy, the spread of predicted classes across rewordings of the same intent is governed by the soundness of the model's parametric knowledge (conceptual certainty) and by lexical variation in the prompt. The paper tests this by generating 15 paraphrase variants for each latent intent, running greedy inference on each variant, and using the highest vote frequency as the ensemble accuracy. Over 179 labeled questions, the ensemble-accuracy distributions for correct and incorrect predictions differ strongly (KS statistic $0.809$, $p = 7.143\times10^{-9}$), which is the empirical basis for using the two empirical cumulative distribution functions (ECDFs) to estimate the probability that a new prediction is correct.

Load-bearing premise

The entire uncertainty estimate depends on the 15 automatically generated rewordings of a question meaning the same thing as the original; if a rephrasing shifts the meaning, the ensemble no longer measures uncertainty about the intended question.

Editorial extensions

If this is right

  • On the endpoint detector, the fine-tuned model's ensemble_true_label_accuracy rises relative to the base model for several questions (one rates question goes from 0 to 53.33%), which the paper reads as evidence that fine-tuning improves conceptual certainty and reduces sensitivity to lexical variation.
  • At inference time, computing the majority-vote frequency across the 15 paraphrases gives a certainty score that can flag low-certainty predictions before the model's answer is acted on.
  • Because the two ensemble-accuracy distributions are separated (Kolmogorov-Smirnov statistic 0.809), the ECDF tail probabilities give a concrete estimate of whether a new prediction is correct.
  • The paper notes that class-specific distributions of ensemble accuracy would give more accurate probability estimates than the global distributions used here, but would require substantially more labeled samples per class.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper stops at using the two ECDFs as a relative score; a natural next step is calibration, where the majority-vote frequency is binned and each bin's observed accuracy is compared with the ECDF-derived probability to see whether the score is a true probability rather than just a ranking.
  • Because the method only requires rephrased inputs and the model's votes, it should transfer to other structured outputs such as tool calls or multi-label extractions; the paper's parameter detector already points in that direction, where sufficiency rather than the single most frequent class is the harder target.
  • Using a different model to generate the paraphrases than the model being evaluated would test how much of the measured variance is a paraphrase artifact and how much is the classifier's own conceptual certainty.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The manuscript proposes that, under greedy decoding, the variance of an LLM's classification output is a function of a 'conceptual certainty' embedded in the model's parametric knowledge and of the lexical variance of the input, and that fine-tuning reduces sensitivity to lexical variation. It operationalizes this via an ensemble of 15 automatically generated paraphrases per question, defines ensemble accuracy as the majority-vote share, and builds empirical CDFs for correct and incorrect predictions on 179 labeled endpoint-detection questions. The authors report that the two CDFs differ according to a Kolmogorov-Smirnov test and use the CDFs to estimate whether a new prediction is correct. The paper also presents bar charts comparing base and fine-tuned models on endpoint and parameter detectors.

Significance. The idea of using paraphrase ensembles to quantify classification uncertainty is practically relevant, and the paper's framing as a black-box method is attractive. The manuscript is honest about the small size of the incorrect-prediction sample and about the risk of paraphrase intent drift. However, as it stands the central variance decomposition (Eq. 2) is an untested hypothesis with an unspecified function F and no operational measure of conceptual certainty, and the proposed probabilistic calibration is evaluated on the same data used to build the distributions. Without a held-out test set and a comparison to existing uncertainty quantification baselines, the 92.7% prediction accuracy and the KS result do not establish predictive value. The paper's contribution is therefore currently a method sketch rather than a validated uncertainty quantification method.

major comments (5)
  1. [Section 2, Eq. (2)] The central variance decomposition is asserted rather than derived: F is never specified, and 'conceptual_certainty' has no operational definition independent of the effects it is invoked to explain. As written, Eq. (2) cannot be tested or falsified; the paper needs either a formal derivation or a concrete experimental protocol that could disconfirm the decomposition.
  2. [Section 4] The evaluation of the probabilistic method is in-sample and circular. The same 179 labeled questions are used to compute ensemble accuracies, construct the two empirical CDFs (ensemble_certainty_correct and ensemble_certainty_incorrect), and then report the KS statistic and the 92.7% prediction accuracy. This only shows that the ECDFs describe the training questions; it does not show that U_new predicts correctness for unseen questions. The authors should hold out a test set (or use cross-validation) and report the distribution of P(correct | U_new) on unseen labels. The illustrative probabilities in Table 5 are not evidence of predictive performance.
  3. [Section 3] The ensemble method assumes that the 15 automatically generated paraphrases preserve the latent intent Q_i. The paper acknowledges that variants may deviate from the original intent, but it provides no independent validation of intent preservation, e.g., human annotation or entailment-based filtering. If a paraphrase changes the intent, the ensemble accuracy measures disagreement about different questions rather than uncertainty about the original question.
  4. [Sections 1 and 4] The proposed method is not compared to any existing uncertainty quantification approach, despite the literature review in Section 1 (semantic entropy, token entropy, self-consistency, Unc-TTP). To support the claim that the ensemble-accuracy CDFs are useful, the authors should benchmark against at least one baseline on the same 179 questions, reporting calibration and accuracy at detecting incorrect predictions.
  5. [Sections 3.1 and 4] The incorrect-prediction distribution is based on a very small number of questions (the paper calls this 'a source of sampling size error'), and the reported KS p-value therefore has low power. More importantly, the two empirical CDFs are built from only 179 questions, and the claimed 92.7% accuracy is not accompanied by confidence intervals or any measure of variability. The authors should report the number of incorrect predictions, bootstrap intervals, and the sensitivity of the ECDF estimates to the choice of n=15.
minor comments (6)
  1. [Section 3] The indexing is inconsistent: 'n=15 variations' is followed by sets {Q_i0,...Q_i15}, which implies 16 variants; please clarify whether n is the number of variants or the maximum index.
  2. [Conclusion] The phrase 'greated conceptual certainty' contains a typo and should read 'greater conceptual certainty'.
  3. [Section 4 and Table 5] The statements such as 'more likely to have come from the correct distribution' conflate the CDF/survival values with posterior probabilities; rephrase as 'the evidence is higher/lower under the two fitted distributions.'
  4. [Section 4] The sentence 'The prediction accuracy of endpoint detection was 92.7%' is introduced abruptly; specify how this accuracy is computed and on which set.
  5. [Section 3] The manuscript does not state how the 179 labeled questions were selected from the space Q_i, nor whether they are independent of the questions used in the detector evaluations in Section 3.1; this should be clarified.
  6. [Tables 2-4] The figure and table captions do not explain how 'accuracy' in the chart titles is computed; the definition of ensemble_true_label_accuracy is given in the text, but a reader looking only at the tables cannot reconstruct the plots.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the in-sample calibration is a validity limitation, not a circular prediction.

full rationale

The paper's derivation chain does not reduce to its inputs by construction. Equation (2) is an informal hypothesis with an unspecified F, so it is unfalsifiable as stated, but it is not used to derive the empirical results. The probabilistic method in Section 4 constructs two empirical CDFs from the 179 labeled questions and then uses those CDFs to evaluate a new ensemble-accuracy value U_new. This is a calibration procedure, not a circular prediction: the ECDFs are fitted to labeled data, and evaluating them at a new point is a standard nonparametric estimate, not an equivalence between the prediction and the fit. The paper does not claim to validate the calibration on the same 179 questions; the 92.7% figure is the ensemble majority-vote accuracy on the 179 questions, and the KS statistic (0.809, p = 7.143e-09) is a descriptive comparison of the two in-sample distributions. Table 5 shows illustrative CDF lookups, not an in-sample accuracy of the correctness estimator. The lack of a held-out test set is a genuine external-validity concern, and the paraphrases' intent preservation is an unvalidated assumption, but neither makes the derivation circular. No load-bearing self-citations, imported uniqueness theorems, or ansatz-by-citation steps appear. The paper's central estimates are empirical fits, and the paper is honest that the incorrect-prediction distribution is small; the weakness is statistical evidence, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claims rest on a posited variance decomposition (Eq. 2), an unvalidated assumption that paraphrases preserve intent, and empirical calibration using fitted ECDFs from the same 179 labeled questions. The only invented entity is the latent construct 'conceptual_certainty,' which has no independent evidence.

free parameters (3)
  • number of paraphrase variants n = 15
    Chosen by hand; the ensemble size directly determines the resolution of the certainty estimate and the variance of the vote distribution.
  • empirical CDF of ensemble_certainty_correct = ECDF computed from correctly predicted questions in the 179-item labeled set
    Fitted to training data; used to compute P_correct(U < U_new) for new predictions.
  • empirical CDF of ensemble_certainty_incorrect = ECDF computed from incorrectly predicted questions in the 179-item labeled set
    Fitted to training data; used to compute P_incorrect(U > U_new) for new predictions.
assumptions (4)
  • ad hoc to paper LLM output variance decomposes as Eq. (2) into conceptual certainty and input variance.
    This is the paper's central hypothesis, introduced in Section 2 as a 'posit' without derivation or external justification.
  • domain assumption Fine-tuning improves conceptual certainty and thereby reduces sensitivity to lexical input variation.
    Assumed and tested on a single endpoint detection task; no theoretical argument or cross-task confirmation is offered.
  • domain assumption Majority vote over the paraphrase ensemble is a valid prediction rule.
    The paper uses the highest-frequency class as the prediction without comparing to other aggregation strategies or justifying it statistically.
  • standard math The KS test assumptions of independence and identical distribution hold for the two certainty distributions.
    Applied in Section 4 without checking sample overlap, autocorrelation from repeated paraphrases of the same intent, or adequacy of the small incorrect class sample.
invented entities (1)
  • conceptual_certainty
    purpose: A latent construct hypothesized to govern the sensitivity of LLM outputs to lexical input variation
    No direct measurement is provided; it is only proxied by ensemble accuracy on one task, so there is no falsifiable handle outside this paper's setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ensemble based approach to quantifying uncertainty of LLM based classifications." pith.science (2026). https://pith.science/paper/K7YHUWIT

@misc{pith2026250208631,
  author       = {Pith},
  title        = {Pith review of: Ensemble based approach to quantifying uncertainty of LLM based classifications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7YHUWIT}},
  note         = {Machine review of arXiv:2502.08631}
}
read the original abstract

The output of Large Language Models (LLMs) are a function of the internal model's parameters and the input provided into the context window. The hypothesis presented here is that under a greedy sampling strategy the variance in the LLM's output is a function of the conceptual certainty embedded in the model's parametric knowledge, as well as the lexical variance in the input. Finetuning the model results in reducing the sensitivity of the model output to the lexical input variations. This is then applied to a classification problem and a probabilistic method is proposed for estimating the certainties of the predicted classes.

Figures

Figures reproduced from arXiv: 2502.08631 by the authors.

Figure 1
Figure 1. Concept space In short the above can be summarized as the following: Finetuning -> Improves "conceptual certainty" -> var(LLM classification) becomes less dependent on var(LLM input) 2.1 How does it help with training? From a concept graph perspective, it identifies the need to strengthen disambiguation either through definition or examples. Responses fall into the following three categories: 1. Confident (i.e., mor… view at source ↗
Figure 2
Figure 2. Distribution of certainties for correct and incorrect predictions [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 7 canonical work pages

  1. [1]

    Llm lies: Hallucinations are not bugs, but features as adversarial examples.,

    J.-Y. Yao, K.-P. Ning, Z.-H. Liu, M.-N. Ning, Y.-Y. Liu, and L. Yuan, “Llm lies: Hallucinations are not bugs, but features as adversarial examples.,” arXiv preprint arXiv:2310.01469, 2023. 10

  2. [2]

    LM-Polygraph: Uncertainty Estimation for Language Models,

    E. Fadeeva et al., “LM-Polygraph: Uncertainty Estimation for Language Models,” 2023

  3. [3]

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation,

    L. Kuhn, Y. Gal, and S. Farquhar, “Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation,” ICLR, 2023

  4. [4]

    Unsupervised quality estimation for neural machine translation. ,

    M. Fomicheva et al., “Unsupervised quality estimation for neural machine translation. ,” Transactions of the Association for Computational Linguistics, 8:539–555, 2020

  5. [5]

    Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models.,

    J. Duan et al. , “Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models.,” Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

  6. [6]

    Generating with confidence: Uncertainty quantification for black-box large language models.,

    Z. Lin, S. Trivedi, and J. Sun, “Generating with confidence: Uncertainty quantification for black-box large language models.,” CoRR, 2023

  7. [7]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models ,

    X. Wang et al. , “Self-Consistency Improves Chain of Thought Reasoning in Language Models ,” arXiv:2203.11171

  8. [8]

    Unc-TTP: A Method for Classifying LLM Uncertainty to Improve In-Context Example Selection

    H.-Y. Huang, Z. Wu, Y. Yang, J. Zhang, and Y. Wu, “Unc-TTP: A Method for Classifying LLM Uncertainty to Improve In-Context Example Selection.”

Show all 9 references
  1. [9]

    Detecting hallucinations in large language models using semantic entropy,

    S. Farquhar, J. Kossen, L. Kuhn, and Y. Gal, “Detecting hallucinations in large language models using semantic entropy,” Nature, 2024. 11

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.