Pith. sign in

REVIEW 2 major objections 5 minor 33 references

With zero-noise unit labels, LLM confidence ranking largely fails at atomic resolution, and reasoning boosts accuracy while harming the ability to rank errors.

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 · grok-4.5

2026-07-11 23:22 UTC pith:6YVIGAUV

load-bearing objection Clean zero-noise long-form uncertainty benchmark plus three solid empirical regularities; main limit is how far synthetic single-answer tasks travel. the 2 major comments →

arxiv 2607.03870 v1 pith:6YVIGAUV submitted 2026-07-04 cs.AI cs.CLcs.LG

Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth

Ido Amit , Ido Galil , Ran El-Yaniv This is my paper
classification cs.AI cs.CLcs.LG
keywords LLM uncertainty estimationlong-form generationdeterministic ground truthatomic evaluationconfidence rankingcalibrationreasoning trade-offprefix error propagation
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.

Long-form LLM outputs contain many local pieces, some right and some wrong, so uncertainty tools must flag the bad pieces rather than reject the whole response. Existing long-form benchmarks introduce label noise through judges or multi-valid answers, which can systematically distort ranking and calibration metrics. This paper introduces SALT, a suite of six procedurally generated tasks (math, code, logic, translation, multi-needle retrieval) that have a single deterministic long textual ground truth, so every atomic unit can be scored by exact string match without any external judge. Across 50+ models the authors show three concrete results: raw confidence signals often fail to separate correct from incorrect atoms even when coarser line-level ranking still works; future correctness is driven by two separable factors—error propagation from corrupted prefixes (global prefix quality dominates local) plus a bounded length-related degradation that saturates early; and both CoT prompting and trained reasoners improve precision while systematically degrading confidence ranking. The practical claim is that risk-critical systems need high-resolution, zero-noise evaluation and must treat reasoning’s accuracy gain as potentially paid for by worse self-ranking of errors.

Core claim

On a deterministic long-form benchmark with exact unit labels, confidence ranking largely collapses at atomic resolution for most models even when line-level ranking remains informative; controlled prefix interventions separate two drivers of later errors—propagation dominated by global context correctness and a bounded length effect—and reasoning (CoT or specialized training) improves accuracy while degrading ranking ability.

What carries the argument

SALT (Single-answer Atomic Long-form Target): six procedurally generated tasks with one known long textual ground truth, enabling exact unit-level correctness labels, multi-granularity scoring, and controlled atom-level prefix interventions without judges or noisy decomposition.

Load-bearing premise

That the error dynamics, atomic ranking failure, and reasoning–ranking trade-off measured on these six single-answer structured tasks with strict index-aligned string matching transfer to open-ended long-form settings where multiple answers can be valid.

What would settle it

Re-run the same models and confidence functions on open-ended long-form tasks with high-quality multi-annotator atomic labels (or another zero-noise multi-valid setting) and check whether atomic AUROC remains near chance, the global-prefix propagation effect replicates, and the reasoning-induced AUROC drop still appears; if atomic ranking becomes strong or the trade-off vanishes, the central claims do not transfer.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces SALT, a procedurally generated long-form benchmark of six single-answer tasks (math, code, logic, DNA translation, multi-needle) with deterministic unit-level ground truth, enabling zero-noise evaluation of precision, ECE, and AUROC at atomic and line resolutions without external judges. Across 50+ LLMs it reports that (i) logits-based confidence functions dominate verbalized ones, with different functions best for calibration vs ranking; (ii) confidence ranking largely fails at atomic resolution while remaining informative at line level; (iii) controlled prefix interventions (mainly DNA) separate two drivers of future error—propagation from corrupted prefixes, with global correctness dominating local, and a bounded length-related degradation; and (iv) reasoning (CoT or trained reasoners) improves precision while systematically degrading AUROC. Code is released.

Significance. If the within-SALT results hold, the paper supplies a clean, contamination-resistant substrate for fine-grained uncertainty evaluation that prior long-form benchmarks lack because of judge noise. The atomic ranking failure, separable prefix vs length drivers, and reasoning–ranking trade-off are concrete, falsifiable findings with direct implications for selective generation and risk-critical deployment. Strengths include the large multi-model sweep, Wilcoxon dominance tests, mediation/ANM-style analysis of precision→ECE, and controlled interventions with mutual adjustment and saturation diagnostics (Appendix B). External representativeness remains the main limit on impact, which the paper itself flags via conditional AIME/MMLU-Pro transfer.

major comments (2)
  1. Section 5.1 and Appendix B.2: the causal claim that future correctness has two separable drivers (global prefix correctness dominating local, plus bounded length degradation) rests primarily on DNA interventions chosen for low inter-atom semantic dependence. Appendix K formalizes that other tasks (logic, multi-needle, matrix mult) have stronger input-span or logical-output dependence. Without analogous interventions on at least one higher-dependence task, the generality of the two-driver claim beyond DNA is under-supported for the main-text framing.
  2. Section 5.2 (reasoning trade-off) and Appendix G.5–G.6: the AUROC degradation under CoT/reasoning is a central claim, but the paired Instruct vs Reasoning comparisons and hybrid CoT+reasoning analysis are reported mainly in aggregate. Task-level and model-pair breakdowns (Figures 39–40) show substantial heterogeneity; the manuscript should state more clearly for which tasks/models the ranking degradation is robust versus precision-driven or task-specific, so the trade-off is not over-generalized from the median effect.
minor comments (5)
  1. Section 4.2 vs abstract/intro: the text mentions eight tasks then six with full coverage (ARC-AGI and Maze held out). Align the abstract and contribution list with the six-task main aggregate to avoid confusion.
  2. Figure 1 and Section 2.3: ECE and AUROC are illustrated at generation/line/atom levels; a short note that generation-level AUROC is often undefined or degenerate when entire generations are all-correct or all-incorrect would help readers interpret the granularity gap.
  3. Appendix H: the Needleman–Wunsch alignment ablation is useful; a one-sentence pointer in Section 4.3 to why strict indexing is preferred for precision (redundant units) would strengthen the main-text justification.
  4. Table 1 / task sizes: DNA contributes ~29k of ~55k atoms. Confirm that task-equal averaging (stated in Appendix A) is used for all main figures so DNA does not dominate aggregate AUROC/ECE.
  5. Typos and polish: e.g., 'Words Collection' prompt figure caption reused for Kronecker in one place; 'Maro-PRR' in a figure caption; minor notation consistency for U_gen vs U_gt.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark with independent deterministic labels, standard metrics, and external consistency checks.

full rationale

SALT is a procedurally generated benchmark whose ground-truth unit sequences are known a priori and independent of any model confidence scores; correctness is strict index-aligned string equality after fenced extraction. Precision, ECE, and AUROC are standard metrics applied to those labels. Confidence functions (perplexity, entropy, logprobs sum, verbalized probes) and post-hoc calibrations (Z-sigmoid, MinMax, binned) are compared empirically; selecting the best-performing function/scheme per metric for reporting is ordinary model-selection practice, not a fitted parameter renamed as a prediction of the same quantity. Prefix interventions are controlled counterfactuals on the model’s own answer context (primarily DNA), with paired baselines and spline regressions; they do not define the outcome by construction. Reasoning trade-off results come from paired Instruct/Reasoning and CoT comparisons plus mediation (DoWhy/ANM), not from self-definition. Self-citations (e.g., Galil/El-Yaniv selective-prediction background, Goren et al. 2026 motivational) are non-load-bearing. Consistency with MMLU-Pro, GPQA, Arena, and conditional transfer to AIME further anchors results outside the paper’s own fits. No self-definitional loop, no uniqueness theorem imported from the authors, no ansatz smuggled via self-citation, and no renaming of a known pattern as a first-principles derivation. The paper is self-contained empirical measurement; circularity score is zero.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing content is methodological: correctness is exact match on procedurally unique targets; uncertainty is read from token logits or verbalized probes; evaluation units are task-defined spans. No physical constants or fitted universal laws. Free parameters are standard evaluation hyperparameters (ECE bins, spline df, calibration schemes). Invented constructs are the SALT task suite and the formal evaluation-unit partition used for labeling.

free parameters (5)
  • ECE number of bins m = 15
    Set to 15 following Guo et al.; changes binning and thus reported ECE.
  • Binned calibration bin counts B = B in {5,10,100}
    Evaluated at 5, 10, 100; best scheme chosen per metric, affecting ECE vs AUROC trade-off.
  • Cubic B-spline degrees of freedom for intervention curves = df=5
    df=5, deg=3 with explicit boundary knots; shapes estimated prefix and length effects.
  • Problem size L and matrix dimension constraints = task-dependent L
    Controls instance difficulty and atom counts for MatMul/Kronecker (Table 3); chosen by design, not learned from model errors.
  • Precision regime split threshold 0.6 for causal analysis = 0.6
    Used to stratify non-monotonic Precision→ECE relationship in Appendix G.5; affects mediation conclusions.
axioms (5)
  • domain assumption A generated unit is correct iff it exactly matches the corresponding ground-truth unit string (strict index alignment by default).
    Section 2.1 and 4.3; defines all precision/AUROC labels. Alignment alternative explored only in Appendix H.
  • domain assumption Token-level model probabilities (or verbalized probes) can be aggregated into unit-level confidence scores that are meaningful for calibration and ranking.
    Section 2.2 and Table 7; standard in information-based uncertainty estimation.
  • ad hoc to paper Procedurally generated single-answer structured tasks are a valid substrate for studying long-form uncertainty dynamics relevant to deployment.
    Design objectives in Section 4.1; transferability tested but only partially supported in Section 5.2.
  • standard math AUROC under 0/1 unit correctness equals the probability that a correct unit outranks an incorrect one (ranking risk specialization).
    Section 2.3 citing Tortorella / ranking risk literature.
  • domain assumption Final-answer fencing isolates the evaluated sequence from reasoning tokens without changing content semantics.
    Section 2 and 4.3 response processing pipeline.
invented entities (3)
  • SALT (Single-answer Atomic Long-form Target) benchmark suite independent evidence
    purpose: Provide unbounded, contamination-resistant long-form instances with unique deterministic ground truth and task-aligned units.
    Core contribution of the paper; independent evidence is the public code and procedural generators, not external physical measurement.
  • Evaluation units (atomic-unit / line-unit partition of a generation) no independent evidence
    purpose: Define intermediate granularity for correctness and confidence aggregation between tokens and full generations.
    Formalized in Section 2.1; task-specific delimiters implement the partition.
  • Redundant units (task-specific intrinsic hallucination subtype) no independent evidence
    purpose: Label generated units absent from the ground-truth sequence as a localized hallucination form.
    Section 2.1 definition used for precision sensitivity to extras.

pith-pipeline@v1.1.0-grok45 · 56662 in / 3687 out tokens · 38878 ms · 2026-07-11T23:22:07.542388+00:00 · methodology

0 comments
read the original abstract

As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses. While such methods exist, evaluating uncertainty at any resolution (token to an entire generation) is challenging and highly sensitive to label imperfections, making zero-noise benchmarks essential; yet, long-form generation benchmarks tend to rely on fallible labels rather than deterministic ground truth. We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges. Equipped with SALT, our analysis of 50+ LLMs reveals key insights: We identify which confidence functions dominate each uncertainty aspect and show that confidence ranking largely breaks at atomic resolution, even when clearer separability emerges at coarser line-level units. SALT further enables controlled atom-level interventions throughout generation, revealing two separable drivers of future errors: propagation from corrupted prefixes, dominated by global context correctness, and bounded degradation from increasing answer-context length. Finally, we demonstrate that reasoning, via Chain-of-Thought prompting or internalized through training, introduces a trade-off, improving accuracy while degrading confidence ranking. These findings directly impact risk-critical applications requiring reliable error identification and mitigation.

Figures

Figures reproduced from arXiv: 2607.03870 by Ido Amit, Ido Galil, Ran El-Yaniv.

Figure 1
Figure 1. Figure 1: Granular uncertainty evaluation with SALT. Decom￾posing long-form generation into high-resolution units isolates localized errors, revealing precise calibration (ECE) and ranking (AUROC) signals obscured by coarse, generation-level evaluation. 1. Introduction Large Language Models (LLMs) have demonstrated remark￾able capabilities across a diverse array of Natural Language Processing (NLP) tasks, increasing… view at source ↗
Figure 2
Figure 2. Figure 2: Probability density of atom correctness relative to its position within the generation. The x-axis represents normalized length (e.g., x = 0.6 is the 60th atom in a 100-atom sequence). How Context Correctness and Answer Length Shape Future Correctness. To test whether the correctness of the generated prefix itself affects future correctness, we perform a controlled counterfactual intervention on the model’… view at source ↗
Figure 3
Figure 3. Figure 3: Probability of atom correctness (y-axis) vs. proportion of prior correct atoms (x-axis). Points represent bins of atoms grouped by their correctness history; markers show bin means, and error bars indicate standard deviation. (last-10 atoms) correctness. Answer-context length is not in￾tervened on directly; instead, we estimate its adjusted associ￾ation with future correctness by evaluating length-response… view at source ↗
Figure 6
Figure 6. Figure 6: Impact of reasoning methods on Precision, ECE, and AUROC. Each marker represents the relative change between two identical models, differing only by CoT prompting or reasoning￾specific training. Markers above the zero line represent an im￾provement in that metric relative to the base model’s performance, while markers below represent a degradation. in uncertainty estimation is whether a single confidence f… view at source ↗
Figure 5
Figure 5. Figure 5: Macro-AUROC at atomic-unit and line-unit resolutions. time reasoning, and models trained specifically to reason (Jaech et al., 2024; DeepSeek-AI, 2025). To isolate these effects, we compare paired Instruct and Reasoning models of identical architecture and size, and applying CoT only to the Instruct variants. Our results reveal a reasoning trade￾off: a systematic inverse relationship between accuracy and r… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of post-hoc calibration schemes. Bold labels indicate the best-performing method, with the corresponding opti￾mal confidence function listed below. Purple and green markers denote atomic- and line-unit granularities, respectively. well to AIME at line-level resolution, while MMLU-Pro agreement remains weak. See Appendix G.3 for details. 6. Conclusion In this work, we introduced SALT, a long-form… view at source ↗
Figure 7
Figure 7. Figure 7: Statistical dominance of confidence functions. Green cells denote that the row function significantly outperforms the column function (p < 0.05, one-sided Wilcoxon signed-rank test). Evaluations are shown for the atomic-unit resolution. reported uncertainty analyses use the best-performing cali￾bration method for each metric, ensuring fair comparisons. 5) Uncertainty Estimation Transferability. Finally, we… view at source ↗
Figure 9
Figure 9. Figure 9: The Matrix Multiplication task prompt, demonstrating few-shot examples and strict output fencing instructions. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The Translation task prompt, demonstrating few-shot examples and strict output fencing instructions. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The Kronecker Product prompt, demonstrating few-shot examples and strict output fencing instructions. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The First-Order Logic task prompt, demonstrating few-shot examples and strict output fencing instructions. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The Code task prompt, demonstrating few-shot examples and strict output fencing instructions. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: The Words Collection task prompt, demonstrating few-shot examples and strict output fencing instructions [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of precision across tasks at atomic and line-unit resolutions. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of ECE across tasks at atomic and line-unit resolutions. Confidence is computed using Perplexity, the optimal function for calibration established in Section 5.2 [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: The measured relative redundant atoms in generations, across tasks. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: ARC-AGI precision at atomic and generation-level resolutions. Atom-level evaluation reveals high partial cell-wise correctness that is obscured by poor exact generation-level scoring. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: ARC-AGI Macro-AUROC comparison at atomic with Line and generation resolutions. Both coarse resolutions often provide a clearer confidence ranking signal than atom-level ranking. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Maze precision at atomic and line-level resolutions. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.2 0.25 0.3 0.35 0.4 Maze: Precision vs Macro-AUROC - Atom granularity Macro-AUROC P recision Fit R²=0.97 ρ=-0.80 Random Guess [PITH_FULL_IMAGE:figures/full_fig_p027_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Maze atom-level precision versus Macro-AUROC. The evaluated models show a strong negative association, suggesting that higher path accuracy does not necessarily imply better confidence ranking, and vice versa. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: A comparison across the benchmark’s tasks, of the probability of an atom being correct given the percentage or prior correct atoms. Second, we validate whether different prompting method plays a role, where we compare a regular prompt with a Chain-of￾Thought prompt [PITH_FULL_IMAGE:figures/full_fig_p028_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: A comparison between regular and Chain-of-Thought prompting, of the probability of an atom being correct given the percentage or prior correct atoms. B.2.1. PROTOCOL AND TREATMENTS Let a (p) m,s denote the atom at position p in model m’s original generation on sample s. For each triple (m, s, p⋆ ) we retain the original prefix a (1) m,s, . . . , a (p ⋆−1) m,s and produce a family of modified prefixes by d… view at source ↗
Figure 24
Figure 24. Figure 24: A comparison between instruction-tuned and reasoning models, of the probability of an atom being correct given the percentage or prior correct atoms. equals 0 on baseline rows by construction, anchoring the regression at the no-intervention point. Pairing absorbs all (m, s, p⋆ )-level fixed effects — baseline difficulty, sample-specific lexical context, model-specific generation tendencies — at a granular… view at source ↗
Figure 25
Figure 25. Figure 25: Effect of controlled prefix-correctness perturbations on future correctness. Length-response curve [PITH_FULL_IMAGE:figures/full_fig_p032_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Length-response curves, controlled vs. uncontrolled for prefix correctness. Both saturate on the same timeline; the uncontrolled curve drops to a deeper asymptote, attributable to prefix-correctness confounding. Legend reports the joint-significance p-value, the slope-knee with 95% CI, and the 95%-asymptote position with 95% CI for each curve. Fixed-correctness curves. Holding prefix correctness fixed at … view at source ↗
Figure 27
Figure 27. Figure 27: Predicted P(next correct) as a function of preceding answer atoms with prefix correctness held fixed at different levels [PITH_FULL_IMAGE:figures/full_fig_p034_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Effect of post-hoc calibration’s schemes on ranking. We observed that the best-performing confidence functions on the ECE metrics consistently were among those that require normalization. One can argue that the calibration step unfolds the probability distribution in a way that ECE benefits from it. To test that possibility, we experimented with calibrating all confidence functions’ scores, regardless of … view at source ↗
Figure 29
Figure 29. Figure 29: ECE (y-axis) as generation unfolds (x-axis). Points represent bins of atoms’ relative position. Markers show bin means. value signifies higher confidence (↑) or higher uncertainty (↓). We additionally evaluate verbalized confidence functions, which elicit a confidence value by querying the model itself rather than aggregating token-level statistics. Two probe types are considered: P(True) (Kadavath et al.… view at source ↗
Figure 30
Figure 30. Figure 30: Which confidence function is optimal for Calibration among the tested confidence functions. 48 [PITH_FULL_IMAGE:figures/full_fig_p048_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Which confidence function is optimal for Ranking among the tested confidence functions. 49 [PITH_FULL_IMAGE:figures/full_fig_p049_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Which confidence function is optimal for the prediction-rejection ratio among the tested confidence functions. 50 [PITH_FULL_IMAGE:figures/full_fig_p050_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Relationship between LLM precision and confidence ranking metrics. 51 [PITH_FULL_IMAGE:figures/full_fig_p051_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Per-model rank correlation 52 [PITH_FULL_IMAGE:figures/full_fig_p052_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Average-score rank correlation. 53 [PITH_FULL_IMAGE:figures/full_fig_p053_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: Macro-AUROC and Maro-PRR nearly perfect correlations at the atomic and line-unit levels. 54 [PITH_FULL_IMAGE:figures/full_fig_p054_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: Spearman correlation matrix between PRR and AUROC. Both atom and line units show a high correlation between the two [PITH_FULL_IMAGE:figures/full_fig_p055_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: A comparison of the effect of the CoT and Reasoning combination with each reasoning strategy individually, on Precision, ECE, and AUROC 55 [PITH_FULL_IMAGE:figures/full_fig_p055_38.png] view at source ↗
Figure 39
Figure 39. Figure 39: Impact of reasoning methods on model performance. Each marker represents the relative change between two identical models, differing only by CoT prompting or reasoning-specific training. Markers above the zero line represent an improvement in that metric relative to the base model’s performance, while markers below represent a degradation. 56 [PITH_FULL_IMAGE:figures/full_fig_p056_39.png] view at source ↗
Figure 40
Figure 40. Figure 40: Impact of reasoning methods on model performance. Each marker represents the relative change between two identical models, differing only by CoT prompting or reasoning-specific training. Markers above the zero line represent an improvement in that metric relative to the base model’s performance, while markers below represent a degradation. 57 [PITH_FULL_IMAGE:figures/full_fig_p057_40.png] view at source ↗
Figure 41
Figure 41. Figure 41: Improvement in precision induced by aligning the generated atomic-units with the ground-truth atomic-units using the Needleman–Wunsch algorithm, across tasks [PITH_FULL_IMAGE:figures/full_fig_p058_41.png] view at source ↗
Figure 42
Figure 42. Figure 42: Simulation of AUROC degradation under varying label noise. The results demonstrate a ”ceiling effect”: as model quality (A) increases, the sensitivity to noise grows non-linearly. This noise acts as a compressor, artificially reducing the gap between top-tier models and random baselines, which masks true progress in uncertainty estimation. 58 [PITH_FULL_IMAGE:figures/full_fig_p058_42.png] view at source ↗
Figure 43
Figure 43. Figure 43: A comparison of the evaluated ECE at the atomic-unit and the line-unit resolutions. (a) PDFs of correct and incorrect atoms’ confidence within the Matrix Multiplication Task (b) PDFs of correct and incorrect atoms’ confidence within the DNA Task [PITH_FULL_IMAGE:figures/full_fig_p059_43.png] view at source ↗
Figure 44
Figure 44. Figure 44: probability density function of z-sigmoid calibrated maximum-probability confidence scores for correct and incorrect atoms on the Matrix Multiplication (44a) and DNA (44b) tasks. The substantial overlap between the two distributions across models illustrates weak atomic-level separability, consistent with the observed high-resolution ranking gap. 59 [PITH_FULL_IMAGE:figures/full_fig_p059_44.png] view at source ↗
Figure 45
Figure 45. Figure 45: Task-level view of the gap between atomic- and line-level ranking. The difference is smaller in DNA Translation than in Multi-Needle, suggesting that the atom-vs.-line ranking gap depends on task structure rather than arising uniformly across tasks. 60 [PITH_FULL_IMAGE:figures/full_fig_p060_45.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    URL https://aclanthology.org/P04-3031/

    Association for Computational Linguistics. URL https://aclanthology.org/P04-3031/. Bisk, Y ., Zellers, R., Bras, R. L., Gao, J., and Choi, Y . PIQA: reasoning about physical commonsense in nat- ural language. InThe Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conf...

    2020

  2. [2]

    doi: 10.1609/AAAI.V34I05

    AAAI Press, 2020. doi: 10.1609/AAAI.V34I05

    doi:10.1609/aaai.v34i05 2020

  3. [3]

    v34i05.6239

    URL https://doi.org/10.1609/aaai. v34i05.6239. Bl¨obaum, P., G ¨otz, P., Budhathoki, K., Mastakouri, A. A., and Janzing, D. Dowhy-gcm: An extension of dowhy for causal inference in graphical causal mod- els.Journal of Machine Learning Research, 25(147):1– 7, 2024. URL http://jmlr.org/papers/v25/ 22-1258.html. Chen, C., Liu, K., Chen, Z., Gu, Y ., Wu, Y .,...

    doi:10.1609/aaai 2024

  4. [4]

    Chen, L., de Melo, G., Suchanek, F

    URL https://openreview.net/forum? id=Zj12nzlQbz. Chen, L., de Melo, G., Suchanek, F. M., and Varoquaux, G. Query-level uncertainty in large language models, 2026. URLhttps://arxiv.org/abs/2506.09669. Chernozhukov, V ., Chetverikov, D., Demirer, M., Du- flo, E., Hansen, C., Newey, W. K., and Robins, J. M. Double/debiased machine learning for treat- ment an...

    arXiv 2026

  5. [5]

    org/CorpusID:85528548

    URL https://api.semanticscholar. org/CorpusID:85528548. Chow, C. K. An optimum character recognition system using decision functions.IRE Trans. Electron. Com- put., 6(4):247–254, 1957. doi: 10.1109/TEC.1957. 5222035. URL https://doi.org/10.1109/TEC. 1957.5222035. Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. T...

    work page doi:10.1109/tec.1957 1957

  6. [6]

    org/CorpusID:52816654

    URL https://api.semanticscholar. org/CorpusID:52816654. Dawid, A. P. The well-calibrated bayesian.Journal of the American Statistical Association, 77(379):605–610,

  7. [7]

    URL http://www

    ISSN 01621459, 1537274X. URL http://www. jstor.org/stable/2287720. 10 Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth de Boor, C.A Practical Guide to Splines. Applied Math- ematical Sciences. Springer, 1978. ISBN 978-1-4612- 6333-3. doi: 10.1007/978-1-4612-6333-3. URL https: //doi.org/10.1007/978-1-4612-6333-3. DeepSeek...

    doi:10.1007/978-1-4612-6333-3 1978

  8. [8]

    Guo, C., Pleiss, G., Sun, Y ., and Weinberger, K

    URL https://jmlr.org/papers/v6/ gretton05a.html. Guo, C., Pleiss, G., Sun, Y ., and Weinberger, K. Q. On cali- bration of modern neural networks. In Precup, D. and Teh, Y . W. (eds.),Proceedings of the 34th International Con- ference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 ofProceedings of Machine Learning Resea...

    2017

  9. [9]

    Hong, K., Troynikov, A., and Huber, J

    URL http://proceedings.mlr.press/ v70/guo17a.html. Hong, K., Troynikov, A., and Huber, J. Context rot: How increasing input tokens impacts llm performance.URL https://research. trychroma. com/context-rot, retrieved October, 20:2025, 2025. Hoyer, P., Janzing, D., Mooij, J., Peters, J., and Sch¨olkopf, B. Nonlinear causal discovery with additive noise model...

    2025

  10. [10]

    cc/paper_files/paper/2008/file/ f7664060cc52bc6f3d620bcedc94a4b6-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2008/file/ f7664060cc52bc6f3d620bcedc94a4b6-Paper. pdf. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., and Liu, T. A survey on hallucination in large language mod- els: Principles, taxonomy, challenges, and open ques- tions.ACM Trans. Inf. Syst., 43(2):...

    doi:10.1145/3703155 2008

  11. [12]

    org/CorpusID:267200215

    URL https://api.semanticscholar. org/CorpusID:267200215. Malinin, A. and Gales, M. J. F. Uncertainty estimation in autoregressive structured prediction. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net,

    2021

  12. [13]

    Manakul, P., Liusie, A., and Gales, M

    URL https://openreview.net/forum? id=jN5y-zb5Q7m. Manakul, P., Liusie, A., and Gales, M. J. F. Selfcheck- gpt: Zero-resource black-box hallucination detection for generative large language models. In Bouamor, H., Pino, J., and Bali, K. (eds.),Proceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing, EMNLP 2023, Singapore, D...

    doi:10.18653/v1/2023 2023

  13. [14]

    Personalized and Reliable Decision Sets: Enhancing Interpretability in Clinical Decision Support Systems

    URL https://proceedings.mlr.press/ v202/shi23a.html. Tian, K., Mitchell, E., Zhou, A., Sharma, A., Rafailov, R., Yao, H., Finn, C., and Manning, C. D. Just ask for calibra- tion: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. In Bouamor, H., Pino, J., and Bali, K. (eds.),Proceed- ings of the 2023...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/ 2023

  14. [15]

    org/CorpusID:235899031

    URL https://api.semanticscholar. org/CorpusID:235899031. Vashurin, R., Fadeeva, E., Vazhentsev, A., Rvanova, L., Vasilev, D., Tsvigun, A., Petrakov, S., Xing, R., Sadal- lah, A. B., Grishchenkov, K., Panchenko, A., Baldwin, T., Nakov, P., Panov, M., and Shelmanov, A. Bench- marking uncertainty quantification methods for large lan- guage models with lm-pol...

    Pith/arXiv arXiv doi:10.1162/tacl 2025

  15. [16]

    emnlp-main.1543/

    URL https://aclanthology.org/2025. emnlp-main.1543/. Ye, F., Yang, M., Pang, J., Wang, L., Wong, D. F., Yilmaz, E., Shi, S., and Tu, Z. Benchmarking llms via uncertainty quantification. In Globersons, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J. M., and Zhang, C. (eds.),Advances in Neural Information Processing Systems 38: Annual Confere...

    doi:10.18653/v1/2024.acl-long.365 2025

  16. [17]

    emnlp-main.443/

    URL https://aclanthology.org/2024. emnlp-main.443/. Yoon, D., Kim, S., Yang, S., Kim, S., Kim, S., Kim, Y ., Choi, E., Kim, Y ., and Seo, M. Reasoning models better express their confidence.CoRR, abs/2505.14489, 2025. doi: 10.48550/ARXIV .2505.14489. URL https:// doi.org/10.48550/arXiv.2505.14489. Zablotskaia, P., Phan, D., Maynez, J., Narayan, S., Ren, J...

    doi:10.48550/arxiv 2024

  17. [18]

    findings-emnlp.197/

    URL https://aclanthology.org/2023. findings-emnlp.197/. Zadrozny, B. and Elkan, C. Transforming classifier scores into accurate multiclass probability estimates. InPro- ceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 23-26, 2002, Edmonton, Alberta, Canada, pp. 694–

    2023

  18. [19]

    doi: 10.1145/775047.775151

    ACM, 2002. doi: 10.1145/775047.775151. URL https://doi.org/10.1145/775047.775151. Zellers, R., Holtzman, A., Bisk, Y ., Farhadi, A., and Choi, Y . Hellaswag: Can a machine really finish your sentence? In Korhonen, A., Traum, D. R., and M`arquez, L. (eds.), Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Flore...

    doi:10.1145/775047.775151 2002

  19. [20]

    no length

    doi: 10.18653/V1/P19-1472. URL https:// doi.org/10.18653/v1/p19-1472. Zhang, C., Liu, F., Basaldella, M., and Collier, N. LUQ: long-text uncertainty quantification for llms. In Al- Onaizan, Y ., Bansal, M., and Chen, Y . (eds.),Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 12-1...

    Pith/arXiv arXiv doi:10.18653/v1/p19-1472 2024

  20. [21]

    Significance( Hsig 0 :β 1 =· · ·=β 5 = 0): F-test (OLS) / likelihood-ratio test (GLM) against the model with the spline basis removed.Rejects flatness, not linearity

  21. [22]

    correctness

    Nonlinearity( Hnl 0 : the spline model reduces to a linear-in- T model). F-test (OLS) / LR test (GLM) between the spline model and ˜Y=α+δT+γ ⊤X+ε . We also report ∆AIC = AIC(linear)−AIC(spline) ; positive ⇒ spline preferred. Within-triple row correlation makes asymptotic row-level p-values overconfident; we cap reported p-values atp <0.001 and rely on∆AIC...

    2025

  22. [23]

    Regime Stratification:Preliminary non-parametric analysis revealed a non-monotonic ”V-shaped” relationship between precision and calibration. To avoid confounding effects arising from this non-linearity (where opposite trends cancel each 40 Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth Table 9.Causal analysis of the P...

  23. [24]

    Robustness to Confounding (Model Size) Partial Correlation Spearmanρ−0.90(p <0.001)+0.84(p <0.001) Conditional Independence HSIC Test RejectH 0 (p <0.001) RejectH 0 (p <0.001) Causal Effect (DML) Estimate−0.60(p <0.001)+0.62(p <0.001)

  24. [25]

    Causal Directionality (ANM) Forward ModelHSIC p-value0.990.07 (Precision→ECE) DecisionAcceptedAccepted (Ambiguous) Backward ModelHSIC p-value0.030.08 (ECE→Precision) DecisionRejectedAccepted (Ambiguous) other out), we stratified the analysis into two distinct regimes: aLow Precisionregime (Precision <0.6 , N= 34 ) and a High Precisionregime (Precision≥0.6,N= 16)

  25. [26]

    Confounding Assessment (Robustness to Model Scale):Within each regime, we tested whether the Precision-ECE link was spurious and driven simply by model capacity. We controlled forTotal ParametersandActive Parametersusing three complementary tests: •Partial Correlation (Spearman):Measures monotonic association while holding model size constant. • Condition...

    2005

  26. [27]

    We modeled the functional relationship using Gaussian Process regression and tested independence between the predictor and the residuals using HSIC

    Causal Directionality (ANM):To validate the direction of causality, we utilized the Additive Noise Model (ANM) framework (Hoyer et al., 2008). We modeled the functional relationship using Gaussian Process regression and tested independence between the predictor and the residuals using HSIC. A high p-value indicates that the residuals are independent noise...

    2008

  27. [28]

    All three tests (Partial Correlation, HSIC, and DML) confirm that the link persists even after rigorously controlling for model scale

    Robustness:In both regimes, the relationship between Precision and ECE is robust to confounding. All three tests (Partial Correlation, HSIC, and DML) confirm that the link persists even after rigorously controlling for model scale

  28. [29]

    The Low Precision Regime ( <0.6 ):We find a clear causal signal that improvements in precision drive better calibration. The ANM test decisively accepts the forward direction ( p= 0.99 ) and rejects the reverse ( p= 0.03 ), suggesting precision is the functional driver of calibration error in this phase

  29. [30]

    Higher precision causeshighercalibration error, with a positive causal effect of+0.62

    The High Precision Regime (≥0.6 ):The relationship inverts significantly. Higher precision causeshighercalibration error, with a positive causal effect of+0.62. While the directionality tests are ambiguous, we hypothesize it is likely due to the small sample size,N= 16. G.5.3. RESIDUALANALYSIS We fit a linear regression ECE∼Precision on the 35 non-reasoni...

  30. [31]

    Fit: dECE=β 0 +β 1 ·Precision on non-reasoning models

  31. [32]

    For each reasoning modeli, compute residual:r i =ECE i − dECEi

  32. [33]

    Reasoning- Ranking Trade-off

    Test whether{r i}are systematically different from zero using one-sided t-tests Results: • Mean residual:+0.039(slight positive deviation) • One-sided t-test (worse than expected):p= 0.052 • One-sided t-test (better than expected):p= 0.948 The residuals are not significantly different from zero in either direction, indicating reasoning models’ calibration...

    2020

  33. [34]

    On average, this is E[Wk] =n 0A

    Scenario 1 (xk ∈P→N ):We lose the wins xk contributed. On average, this is E[Wk] =n 0A. As a new Negative, xk is compared to remaining Positives. Assuming independence, a random sample is outscored by half the population, so it contributes≈n 1/2losses. 2.Scenario 2 (x k ∈N→P):Symmetric logic applies. The net change in the numerator for a single flip is th...