Fine-Grained Uncertainty Quantification for Long-Form Language Model Outputs: A Comparative Study

David Skarbrevik; Dylan Bouchard; Mohit Singh Chauhan; Viren Bajaj

REVIEW 4 major objections 6 minor 6 references

The simplest uncertainty score — claim-response entailment — outperforms complex alternatives for detecting hallucinations in long-form LLM output.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 22:12 UTC pith:WY52G6L2

load-bearing objection A solid, well-organized comparative study that will be useful to practitioners, but the headline rankings rest on a single unvalidated LLM grader and no significance tests; the core pattern is plausible, yet needs human-validated labels and bootstrap CIs before it should be treated as settled. the 4 major comments →

arxiv 2602.17431 v2 pith:WY52G6L2 submitted 2026-02-19 cs.CL cs.AIcs.LG

Fine-Grained Uncertainty Quantification for Long-Form Language Model Outputs: A Comparative Study

Dylan Bouchard , Mohit Singh Chauhan , Viren Bajaj , David Skarbrevik This is my paper

classification cs.CL cs.AIcs.LG

keywords uncertainty quantificationhallucination detectionlong-form generationclaim-level scoringentailmentuncertainty-aware decodingFactScoretaxonomy

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that, for detecting hallucinations in long-form LLM answers, the straightforward method of checking whether each claim in an answer is entailed by sampled alternative answers is as good as or better than more elaborate methods. Across eight model-dataset combinations, this claim-response entailment score achieved the highest AUPRC in all eight and was within 0.01 AUROC of the best in six, while graph-based, question-answer, and verbalized-confidence scorers rarely beat it. The paper also argues that scoring at the level of atomic claims beats scoring whole sentences, and that uncertainty-aware decoding — dropping low-confidence claims and regenerating the answer from the rest — can raise factual accuracy, in one case from 0.72 to about 0.90. The paper contributes a three-stage taxonomy and a new 400-question STEM/geography long-form QA dataset so methods can be compared fairly.

Core claim

Claim-response entailment is the most reliable unit-level hallucination detector in this study: for every one of the eight LLM-and-dataset combinations, it achieves the best precision-recall performance, and its ROC performance is either the best or within 0.01 of the best. Simple entailment of each atomic claim against full sampled responses beats graph-centrality scorers, question-answer consistency scorers, and verbalized confidence, and it does so at lower computational cost. Claim-level scoring consistently outperforms sentence-level scoring, because sentences bundle multiple facts of mixed reliability. Uncertainty-aware decoding — filtering out low-confidence claims before reconstructi

What carries the argument

The three-stage fine-grained UQ pipeline: decompose a response into units (sentences or claims), score each unit by semantic consistency against sampled responses, and aggregate unit scores into a response-level confidence. Within this pipeline, the key object is the claim-response scorer, which averages an NLI entailment probability between each atomic claim and every sampled full response. The taxonomy distinguishes four scorer families — unit-response, matched-unit, unit-QA, and graph-based — and two aggregation strategies (averaging and uncertainty-aware decoding). The claim-response family is the load-bearing mechanism, and the paper shows it holds up against far more complex alternativ

Load-bearing premise

The ground truth for what is factual comes entirely from one LLM grader (Gemini-2.5-Flash) comparing claims against Wikipedia text, with no human validation; if that grader tends to mark claims as supported when they paraphrase the original response, the relative ranking of scorers could reflect style similarity rather than true factuality.

What would settle it

A human-annotated evaluation on a few hundred long-form QA claims: collect independent human judgments of whether each atomic claim is supported by the cited Wikipedia context, then recompute AUPRC and AUROC for claim-response entailment versus graph-based and QA scorers. If claim-response entailment does not remain at or near the top, the paper's central ranking claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Practitioners building hallucination detectors for long-form output can default to claim-response entailment with about five sampled responses, avoiding the extra cost of graph construction, question generation, or matching.
Uncertainty-aware decoding turns a diagnostic score into a controllable factuality knob: filtering the lowest-confidence claims and regenerating the response reliably improves factual precision, with one demonstrated jump from 0.72 to 0.90 accuracy.
Claim-level decomposition is worth the extra LLM call over sentence-level scoring, because sentence-level detectors cap out far lower (AUROC around 0.69 vs. 0.80).
The new FactScore-STEM-Geo dataset provides a reusable 400-question benchmark spanning four knowledge domains, enabling apples-to-apples comparison of long-form UQ methods.
The taxonomy unifies previously piecemeal methods, making it possible to assemble and compare scorers from interchangeable components rather than treating each method as a bespoke system.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the consistent win of the simplest scorer suggests that the extra complexity in graph-based and QA-based methods may not be buying truthfulness signal — it may be buying structure that is redundant once entailment is measured directly. A testable extension would swap the NLI model for a stronger or weaker one and check whether the ranking of families changes.
The uncertainty-aware decoding result implies a practical deployment pattern: instead of rejecting an entire long-form answer, an end-user system can highlight or delete low-confidence claims, effectively offering a 'lite' version of the answer with guaranteed higher precision — a behavior short-form UQ cannot provide.
The paper's ground-truth protocol uses the same LLM family for grading and for claim decomposition; an obvious high-value extension is human-annotation validation on a subset, since style-similarity bias in the grader could inflate entailment-based scores.
Because the paper fixes a single temperature, NLI model, and embedding model, the robustness of the ranking across those choices is unknown. A follow-up that varies these components would establish whether claim-response entailment remains best across the wider design space the taxonomy defines.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A solid, well-organized comparative study that will be useful to practitioners, but the headline rankings rest on a single unvalidated LLM grader and no significance tests; the core pattern is plausible, yet needs human-validated labels and bootstrap CIs before it should be treated as settled.

read the letter

What you should know first: this is a competent, readable comparative study of fine-grained black-box uncertainty scorers for long-form LLM output. The main finding—claim-response entailment is a strong, cheap default that matches or beats fancier graph and QA-based methods—is consistent across eight model-dataset scenarios and aligns with what I'd expect from prior work. The second finding, that uncertainty-aware decoding can substantially improve factual precision by filtering low-confidence claims, is practically useful and worth taking seriously.

The paper does several things well. The three-stage taxonomy (decomposition, unit scoring, aggregation) is genuinely clarifying and lets the authors present old methods like LUQ, semantic entropy, and graph centrality as instantiations of a common framework. The new FactScore-STEM-Geo dataset, built from long Wikipedia articles across four domains, is a reasonable addition to the evaluation landscape. The ablations on number of sampled responses are also sensible; the diminishing returns beyond m=5 give practitioners a concrete rule of thumb. The code is available, which matters for reproducibility.

The soft spots are not fatal, but they are real. First, the absolute AUROC differences among the top claim-level scorers are often tiny—within 0.01—and the paper gives no confidence intervals or significance tests. For a comparative study whose main message is "the simple method is competitive," that message survives a 0.01 gap, but the claim that entailment is “better or on par” needs some uncertainty quantification attached to the AUROCs themselves. Second, the ground-truth labels come from a single LLM grader (Gemini-2.5-Flash) using the FactScore protocol, and the same model does claim decomposition, question generation, and merging. That is a real methodological vulnerability: if the grader's judgment is biased by how claims are phrased or decomposed, the rankings of scorers could partly reflect the grader's idiosyncrasies. The stress-test note worries this could make the results artifacts. I think "artifact" is too strong a word—the findings are consistent with prior work and with what you'd expect from semantic consistency—but the absence of any human-validated subset or alternative grader is an important gap that should be fixed before the comparative claims are taken as definitive. Third, the UAD results report gains for the top scorer per family selected on the test set; that is a mild selection effect, though the sweeping pattern across thresholds and scenarios mitigates it.

Overall, this is a paper I'd want refereed seriously. It's not a breakthrough, but it's a well-executed synthesis that the hallucination-detection community will cite. I'd recommend a conditional accept: ask for bootstrap CIs on the main AUROC differences, a small human-validated label set or at least a second LLM grader for robustness, and a clearer statement of which scorer was selected for UAD and whether the selection was fixed in advance. For a reading group it's worth a session; for my own work, I'd cite it.

Referee Report

4 major / 6 minor

Summary. The paper proposes a taxonomy for fine-grained black-box uncertainty quantification (UQ) in long-form LLM outputs, organizing methods along three stages: response decomposition, unit-level scoring, and response-level aggregation. Within this framework it formalizes four scorer families — unit-response, matched-unit, unit-QA, and graph-based — and introduces FactScore-STEM-Geo, a new 400-question long-form QA dataset spanning STEM and Geography. Experiments across four LLMs and two datasets evaluate claim- and sentence-level hallucination detection, calibration, response-level scoring, and uncertainty-aware decoding (UAD). The main findings are: (1) claim-response entailment is consistently better than or on par with more complex claim-level scorers, achieving the highest AUPRC in all eight scenarios and AUROC within 0.01 of the best in six; (2) claim-level scoring generally outperforms sentence-level scoring; and (3) UAD filtering improves factual accuracy, e.g., from 0.72 to ~0.90 at the 50th percentile for one model-dataset pair.

Significance. If the empirical findings hold, the paper makes a useful contribution: it unifies several previously disparate methods into a common formal framework, provides open-source code (uqlm), and introduces a new benchmark. The practical recommendation that claim-response entailment is a strong and cheap default is actionable. The clarification of relationships among methods and the explicit comparison of aggregation strategies are valuable for practitioners. The central claims are plausible and largely consistent with prior work. However, the validity of the comparative evaluation rests on an unvalidated LLM-based ground-truth labeler, and the absence of statistical significance testing leaves the central 'better or on par' claims unsupported. The paper is honest about these limitations in §7, but the lack of any robustness check means the headline results may not be reliable.

major comments (4)

[§4.1, §7] Ground-truth labels are produced by Gemini-2.5-Flash using the FactScore protocol, and the same model is used for claim decomposition, claim merging, unit question generation, and grading. No human validation, inter-annotator agreement, or alternative-grader check is reported. This is load-bearing because all AUROC/AUPRC rankings and UAD accuracy gains are measured against these labels. If the grader prefers claims phrased in a way it itself generates, or if it is biased toward consistency across samples (which claim-response entailment directly measures), the superiority of entailment-based scorers and the UAD improvements could be artifacts. The Limitations section concedes that different choices 'could change absolute and relative performance,' but no evidence is provided that the headline patterns survive even a single alternative grader. I recommend evaluating a random subset (e.g.,
[§4.2, Tables 2, 5–8] All AUROC/AUPRC comparisons are point estimates with no confidence intervals or significance tests. Many of the key gaps are smaller than 0.01 (e.g., Table 2: FactScore-Bio Gemini-2.5-Flash, Claim-response 0.794 vs. Graph-based 0.800; FactScore-STEM-Geo Gemini-2.5-Flash, 0.671 vs. 0.673). The central claim that claim-response entailment is 'better or on par with' graph-based scorers cannot be evaluated without knowing whether these differences are within noise. I recommend bootstrap confidence intervals over prompts or a paired significance test (e.g., DeLong for AUROC) for the headline comparisons.
[§4.3] The UAD experiments report only 'accuracy' as a function of filtering threshold, but the metric is not defined. If accuracy is factual precision over retained claims, then increasing the threshold trivially raises precision while reducing recall/coverage. The claim that accuracy rises from 0.72 to ~0.90 at the 50th percentile is not accompanied by recall of retained claims or the response-level FactScore of the reconstructed responses. Without these, the practical benefit of UAD is overstated. I recommend reporting precision-recall curves, the average fraction of claims retained, and the factuality of the final reconstructed response.
[§4.1, §4.2] The comparison between claim-level and sentence-level scoring is confounded by different label sets. For claim-level evaluation, only objective claims are retained, while sentence-level evaluation includes all sentences regardless of objectivity (§4.1). Objective claims are more likely to be verifiable and consistent across samples, which could favor claim-level scorers independent of the granularity effect. The paper's finding (2) ('claim-level scoring generally yields better results than sentence-level') may therefore reflect the labeling criterion rather than the granularity itself. I recommend evaluating sentence-level scorers on the same subset of objective claims (mapping claims back to sentences) or otherwise controlling for label difficulty.

minor comments (6)

[§3.3.4] The formula for betweenness centrality contains ambiguous notation: 'u̸=v̸=s' should be written with explicit conditions 'u≠v, u≠s, v≠s'.
[§3.3.4] The definition of Hmax is ambiguous: 'Hmax = m+ |s|−1 / 2' should be 'Hmax = (m + |s| − 1)/2'.
[§4.1] The paper fixes 'a single temperature' but does not report the temperature or sampling parameters used for response generation. These are critical for consistency-based methods; please add them to the experimental setup.
[Figure 6] The UAD results are shown only as a figure. To support quantitative claims (e.g., '0.72 to ~0.90'), include the underlying values in a table or appendix.
[§4.2, Claims about claim-QA] The paper says manual inspection revealed 'many atomic claims are poorly suited for question-inversion,' but gives no examples. A brief illustration would help readers interpret the poor performance.
[§4.4] Response-level correlations (Tables 3 and 10) are reported only for FactScore-Bio. Reporting the same analysis for FactScore-STEM-Geo would strengthen the generalizability of the response-level claims.

Circularity Check

0 steps flagged

No significant circularity: scorer definitions and empirical evaluations are self-contained against external Wikipedia-grounded FactScore labels; the acknowledged single-grader limitation is a validity concern, not a circular reduction.

full rationale

The paper's scoring formulas (Section 3.3) are defined from semantic consistency functions and sampled responses, with no parameters fitted to the FactScore labels used for evaluation (Section 4.1). The central empirical claims—claim-response entailment's AUPRC/AUROC ranking and UAD accuracy gains—are computed against Wikipedia-grounded FactScore grades, an external benchmark, and the UAD thresholds are percentile-based sweeps (Section 4.3), not optimized fit values. I checked for the enumerated circularity patterns: no self-definitional relation between scorer and label; no fitted input renamed as prediction; the self-citations (uqlm, Bouchard & Chauhan 2025) are implementation details or standard formulas and do not carry the load-bearing comparisons; no uniqueness theorem is imported. The §7 limitation that a single LLM (Gemini-2.5-Flash) was used for claim decomposition, merging, question generation, and grading, and that different choices 'could change absolute and relative performance,' is a real external-validity threat—the skeptical concern about an unvalidated LLM grader—but it is not a circularity: the scorer equations do not reduce to, nor are they defined in terms of, the grading labels. The comparison is therefore independent, though its robustness to grader choice remains untested.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper's empirical findings rest on the validity of LLM-based ground truth, the semantic-consistency proxy, and the reliability of LLM claim decomposition. No fitted free parameters or new theoretical entities are introduced; experimental hyperparameters are fixed but not fitted to maximize performance.

axioms (3)

domain assumption FactScore LLM grading (Gemini-2.5-Flash) against Wikipedia text yields accurate ground-truth factuality labels.
Section 4.1: 'We obtain ground-truth factuality labels using the FactScore grading protocol, which uses an LLM to compare each unit...' No human validation or inter-annotator agreement is reported.
domain assumption Semantic consistency between sampled responses (NLI or embedding similarity) is a valid proxy for factual correctness.
Section 3.2 defines the semantic consistency functions that all scorers rely on; this is the foundational premise of consistency-based UQ.
domain assumption LLM-based claim decomposition, merging, and question generation by Gemini-2.5-Flash produce reliable atomic claims and unique-claim unions.
Section 4.1 uses Gemini-2.5-Flash for decomposition, merging, and question generation; no quality checks or consistency statistics for the decomposition are provided.

pith-pipeline@v1.3.0-alltime-deepseek · 21310 in / 10653 out tokens · 102987 ms · 2026-08-02T22:12:34.709922+00:00 · methodology

0 comments

read the original abstract

Uncertainty quantification has emerged as an effective approach to closed-book hallucination detection for LLMs, but existing methods are largely designed for short-form outputs and do not generalize well to long-form generation. We introduce a taxonomy for fine-grained uncertainty quantification in long-form LLM outputs that distinguishes methods by design choices at three stages: response decomposition, unit-level scoring, and response-level aggregation. We formalize several families of consistency-based black-box scorers, providing generalizations and extensions of existing methods. We also introduce FactScore-STEM-Geo, a new 400-question long-form QA dataset spanning four categories across STEM and Geography. In our experiments across multiple LLMs and datasets, we find 1) claim-response entailment consistently performs better or on par with more complex claim-level scorers, 2) claim-level scoring generally yields better results than sentence-level scoring, and 3) uncertainty-aware decoding is highly effective for improving the factuality of long-form outputs. Our framework clarifies relationships between prior methods, enables apples-to-apples comparisons, and provides practical guidance for selecting components for fine-grained UQ.

Figures

Figures reproduced from arXiv: 2602.17431 by David Skarbrevik, Dylan Bouchard, Mohit Singh Chauhan, Viren Bajaj.

**Figure 2.** Figure 2: Matched-Unit Scoring Workflow 3.3.3 Unit-QA Scorers The unit-qa family of scorers are inspired by the long-form version of semantic entropy proposed by Farquhar et al. (2024). We generalize this approach to allow for alternative consistency functions and granularities. The unit-qa approach relies on a question inversion function as γ : Sg −→ Y that maps a unit (sentence or claim) to a question for which th… view at source ↗

**Figure 3.** Figure 3: Unit-QA Scoring Workflow A unit-QA scorer, cg : Sg × Ym+1 → [0, 1], measures consistency in multiple responses to unit question γ(s): cg(s; y (s) 0 , y (s) cand) = 1 m Xm j=1 η(y (s) 0 , y (s) j ), where, for unit s, y (s) j denotes the jth response to the unit’s question γ(s), g ∈ {sent, claim}, and η ∈ {1 − pc, ˆcos, BertF1,I[y0 = yj ]}. 2 Note that unit-QA effectively applies standard black-box UQ scori… view at source ↗

**Figure 4.** Figure 4: Graph-Based Scoring Workflow Formally, we denote a bipartite graph G with node set V = s ∪ y, where y is a set of m responses generated from the same prompt and s is the union of all unique claims across those decomposed responses. In particular, an edge exists between a claim-response pair (s, y) ∈ s × y if and only if claim s is entailed in response y. We define the following graph metrics for claim s: •… view at source ↗

**Figure 5.** Figure 5: Claim-Level and Sentence-Level AUROC by LLM and Dataset (Higher is Better) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: LLM Accuracy vs. UAD Filtering Threshold by LLM-Dataset (Top per Scorer Type) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: Claim-Response Hallucination Detection AUROC by Number of Sampled Responses [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗

**Figure 8.** Figure 8: Sentence-Response Hallucination Detection AUROC by Number of Sampled Responses [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

**Figure 9.** Figure 9: Matched-Sentence Hallucination Detection AUROC by Number of Sampled Responses [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 10.** Figure 10: Claim-Response Hallucination Detection AUPRC by Number of Sampled Responses [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

**Figure 11.** Figure 11: Sentence-Response Hallucination Detection AUPRC by Number of Sampled Responses [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

**Figure 12.** Figure 12: Matched-Sentence Hallucination Detection AUPRC by Number of Sampled Responses [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗

**Figure 13.** Figure 13: Claim-Level and Sentence-Level AUPRC by LLM and Dataset (Higher is Better) [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗

**Figure 14.** Figure 14: Claim-Level Calibration Plots (Top Scorer per Family) for FactScore-Bio by LLM [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

6 extracted references · 2 linked inside Pith

[1]

Analyze each claim strictly based on the provided context, not your prior knowledge
[2]

14 Xin Liu, Muhammad Khalifa, and Lu Wang

URLhttps://arxiv.org/abs/2407.21057. 14 Xin Liu, Muhammad Khalifa, and Lu Wang. Litcab: Lightweight language model calibration over short- and long-form responses, 2024. URLhttps://arxiv.org/abs/2310.19208. Andrey Malinin and Mark Gales. Uncertainty estimation in autoregressive structured prediction, 2021. URL https://arxiv.org/abs/2002.07650. Potsawee Ma...

Pith/arXiv arXiv 2024
[3]

Yes" or

Respond with "No" if: - The claim contradicts the context - The claim contains information not present in the context - The claim makes assertions that go beyond what the context states Important principles: - Be conservative in your judgments - only mark claims as supported when there is clear evidence - Ignore stylistic differences or paraphrasing if th...

1930
[5]

Respond with "Yes" only if the claim is directly supported by information in the context
[2024]

URLhttps://arxiv.org/abs/2401.01313. Roman Vashurin, Ekaterina Fadeeva, Artem Vazhentsev, Lyudmila Rvanova, Daniil Vasilev, Akim Tsvigun, Sergey Petrakov, Rui Xing, Abdelrahman Sadallah, Kirill Grishchenkov, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, Maxim Panov, and Artem Shelmanov. Benchmarking uncertainty quantification methods for large lang...

Pith/arXiv arXiv 2025
[2025]

URLhttps://openreview.net/forum?id=WOFspd4lq5

ISSN 2835-8856. URLhttps://openreview.net/forum?id=WOFspd4lq5. 13 Dylan Bouchard, Mohit Singh Chauhan, David Skarbrevik, Ho-Kyeong Ra, Viren Bajaj, and Zeya Ahmad. Uqlm: A python package for uncertainty quantification in large language models, 2025. URLhttps: //arxiv.org/abs/2507.06196. Jiuhai Chen and Jonas Mueller. Quantifying uncertainty in answers fro...

arXiv 2025

[1] [1]

Analyze each claim strictly based on the provided context, not your prior knowledge

[2] [2]

14 Xin Liu, Muhammad Khalifa, and Lu Wang

URLhttps://arxiv.org/abs/2407.21057. 14 Xin Liu, Muhammad Khalifa, and Lu Wang. Litcab: Lightweight language model calibration over short- and long-form responses, 2024. URLhttps://arxiv.org/abs/2310.19208. Andrey Malinin and Mark Gales. Uncertainty estimation in autoregressive structured prediction, 2021. URL https://arxiv.org/abs/2002.07650. Potsawee Ma...

Pith/arXiv arXiv 2024

[3] [3]

Yes" or

Respond with "No" if: - The claim contradicts the context - The claim contains information not present in the context - The claim makes assertions that go beyond what the context states Important principles: - Be conservative in your judgments - only mark claims as supported when there is clear evidence - Ignore stylistic differences or paraphrasing if th...

1930

[4] [5]

Respond with "Yes" only if the claim is directly supported by information in the context

[5] [2024]

URLhttps://arxiv.org/abs/2401.01313. Roman Vashurin, Ekaterina Fadeeva, Artem Vazhentsev, Lyudmila Rvanova, Daniil Vasilev, Akim Tsvigun, Sergey Petrakov, Rui Xing, Abdelrahman Sadallah, Kirill Grishchenkov, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, Maxim Panov, and Artem Shelmanov. Benchmarking uncertainty quantification methods for large lang...

Pith/arXiv arXiv 2025

[6] [2025]

URLhttps://openreview.net/forum?id=WOFspd4lq5

ISSN 2835-8856. URLhttps://openreview.net/forum?id=WOFspd4lq5. 13 Dylan Bouchard, Mohit Singh Chauhan, David Skarbrevik, Ho-Kyeong Ra, Viren Bajaj, and Zeya Ahmad. Uqlm: A python package for uncertainty quantification in large language models, 2025. URLhttps: //arxiv.org/abs/2507.06196. Jiuhai Chen and Jonas Mueller. Quantifying uncertainty in answers fro...

arXiv 2025