Pith. sign in

REVIEW 4 major objections 6 minor 17 references

Calibrated confidence is a reliable, cross-model signal that lets a system decide which model to trust, route inputs efficiently, and flag mislabeled data: the paper shows a training-free bin-wise 'confidence advantage' predicts which model

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

2026-08-03 10:56 UTC pith:4CKCEEXT

load-bearing objection Simple, training-free confidence-advantage routing with real empirical coverage, but the central transfer assumption is asserted rather than shown, and one headline result (82% to 90%) strains credibility without released code or error bars. the 4 major comments →

arxiv 2601.07965 v2 pith:4CKCEEXT submitted 2026-01-12 cs.AI cs.LG

When Models Know When They Do Not Know: Calibration, Cascading, and Cleaning

Chenjie Hao , Weyl Lu , Yuko Ishiwaka , Zengyi Li , Weier Wan , Yubei Chen This is my paper
classification cs.AI cs.LG
keywords calibrationconfidencemodel cascadingdata cleaningconfidence advantagemixture-of-expertsvision-languagetraining-free
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 argues that a model's calibrated confidence—the probability it assigns to its own answer—is a reliable and cross-model comparable signal for when the model knows and when it does not. If confidence is calibrated on a validation set, then within any confidence bin the model's accuracy is roughly equal to its average confidence, and this alignment carries over to held-out test data. The paper exploits the fact that two models' calibrated confidences can be compared bin by bin: a 'confidence advantage' computed on validation predicts which model will be more accurate on a new input. That signal powers two training-free applications: cascading a small model with a large one for efficiency with little accuracy loss, and cascading two strong models to exceed both; plus cleaning datasets by flagging samples where multiple confident experts disagree with the given label. The authors evaluate on image classification, code generation, math, reasoning, and knowledge benchmarks. A sympathetic reader would care because it suggests that model uncertainty—when properly calibrated—can be directly reused for routing and quality control without any learned router or retraining.

Core claim

After calibration—temperature scaling for vision, Platt scaling for language—confidence becomes both reliable (monotonic with accuracy) and comparable across models, so that the bin-wise difference in calibrated confidence between two models, computed once on a validation set, predicts which model will be more accurate on a new sample. This 'confidence advantage' is the routing signal for cascading and the confidence-gated disagreement signal for cleaning; no router network or retraining is needed. Empirically, the paper reports that cascading a small and large model keeps accuracy close to the large model while sending a chosen fraction of inputs to the small model, and that cascading two s

What carries the argument

Confidence advantage routing: after calibrating two models on a validation set and partitioning validation samples into equal-count bins by the smaller model's confidence, the paper computes per-bin difference between the large and small model's average calibrated confidence (a^c_i). This bin-wise advantage function is assumed to generalize to new inputs: if a new sample's small-model confidence falls into bin i, its expected accuracy advantage from using the large model is close to a^c_i. The cascade routes to the small model for the K lowest-advantage bins and to the large model otherwise. For cleaning, the corresponding mechanism is cross-model agreement: a label is flagged when all exper

Load-bearing premise

The load-bearing premise is that the confidence-advantage function computed on the validation set remains predictive of the accuracy advantage on new inputs—asserted as a 'generalization property' rather than proved, and showing clear degradation under out-of-distribution corruption.

What would settle it

Measure, on a distribution-shifted set like ImageNet-C at severity 5, whether the validation-derived per-bin confidence advantage a^c_i matches the realized per-bin accuracy advantage a^v_i; if any bin's sign flips and the average performance-gap-recovered drops below the random baseline of 0.5, the routing signal is invalid for that regime.

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

If this is right

  • Cascading with calibrated advantage routing lets a system run a small model on a tunable fraction of inputs while recovering most of a large model's accuracy, without training a router.
  • When the two models are of comparable strength, the same routing can exceed both individual accuracies because each model specializes on parts of the input space.
  • Data cleaning by multi-model high-confidence disagreement recovers more true label errors than the confident-learning baseline at equal precision, on both image and language benchmarks.
  • Calibration learned on a validation split transfers to a held-out test split, so no per-deployment calibration is needed for in-distribution data.
  • The method is insensitive to the specific calibration method and to the number of bins, per the sensitivity analysis.

Where Pith is reading between the lines

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

  • If the generalization property holds broadly, the same bin-wise advantage could be reused as a cheap, no-training router between any pair of models, making adaptive model selection practical at scale.
  • The reported degradation under corruption (APGR falling from 0.80 to 0.54) suggests the advantage signal is distribution-sensitive; a natural extension is to estimate advantage on a small OOD probe set and re-bin at deployment.
  • Because cleaning precision depends on expert independence, ensembling models with different architectures or training data should improve detection power—a testable prediction the paper's framework would support.
  • The cascade's success on same-sized models hints at a general collaborative principle: calibrated confidence can partition a dataset into regions of relative strength without any learned gating, which might extend to more than two models.

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

4 major / 6 minor

Summary. The paper proposes a training-free framework that calibrates confidence for vision and language models, then uses calibrated confidence for two applications: model cascading (routing inputs between models based on a bin-wise 'confidence advantage') and data cleaning (flagging mislabeled samples via multi-model agreement and confidence thresholds). The central methodological assumption is that the confidence advantage computed on a validation set generalizes to new inputs, enabling routing without retraining. Experiments cover ImageNet, MMLU, MBPP, GSM8K, ARC, BigCodeBench, and OOD variants (ImageNet-C, MMLU-Adversarial). The paper reports improved calibration ECE, cascade accuracy-efficiency trade-offs, and data-cleaning precision, including a claim that cascading two ~82% models reaches ~90% accuracy and that cleaned-data cascades reach 96.266% on a cleaned ImageNet subset.

Significance. If the central generalization property held, this would be a simple, broadly applicable alternative to learned routers and would enable practical model collaboration and dataset cleaning. The paper has strengths: a unified calibration formulation, evaluation across multiple domains, comparison to a learned router baseline (RouteLLM), and a human-verified cleaning study for ImageNet. However, the core assumption is asserted without proof, and the paper's own OOD results show the routing signal degrading to near-random under severe corruption (APGR 0.5409 on fog severity 5, close to 0.5). The surprising large-model cascade gains and the cleaned-set accuracy claim also need more scrutiny. The absence of error bars and code/data release further limits verification.

major comments (4)
  1. [Section 2.2, definition of a^c_i] The 'generalization property of the advantage function' is asserted without proof. Calibration generalization for each model's marginal confidence does not imply that the bin-wise difference a^c_i transfers to new data, especially because bins are formed via histogram equalization on the small model's validation confidence. Under distribution shift, a validation bin index no longer indexes the same subpopulation, and per-model calibration can break. The paper's own Table 6 shows APGR dropping to 0.5409 on fog severity 5, i.e., near the random baseline of 0.5, which is exactly the predicted failure mode. Provide a formal statement of the property or a much more extensive empirical validation of bin-wise transfer on OOD data, including per-bin analysis.
  2. [Section 3.3 / Figure 2b] The claim that cascading two medium models (DeiT-B and TinyViT), each around 82% accuracy, reaches 'an astonishing peak accuracy of up to 90%' is extraordinary and lacks supporting detail. No individual model accuracies, per-bin advantage statistics, or oracle upper bound are provided. An 8-point gain over both constituents suggests the routing signal is selecting a very non-representative subset. To make this claim credible, please provide full experimental details, the routing curves, and a sanity check such as an oracle routing upper bound to show the gain is not an artifact of subset selection.
  3. [Table 2 and Section 3.5] The 'Cleaned ImageNet' accuracy of 96.266% is measured on the cleaned validation set (after removing flagged samples), not on the original full validation set. As such, it is not comparable to standard ImageNet accuracy, and the statement that 'the current state-of-the-art models already approach the upper bound of ImageNet performance' is misleading. Report accuracy on the original validation set using the same cleaning protocol (or explicitly state the evaluation denominator), and clarify whether the metric is computed only on the remaining samples.
  4. [Section 3.5 / Table 3 / Appendix B.2] The cleaning precision results are based on only 1,000 manually verified ImageNet samples, and for MMLU they rely on GPT-4o pseudo-labels rather than human annotation. Table 3 is ambiguous: the rows 'Ours 1283 994 0.665' and 'CL 82 0.537' do not clearly define flagged count, overlap, and unique accuracy. Provide explicit definitions, the full annotation counts, and confidence intervals. For MMLU, using GPT-4o as ground truth can bias the evaluation; report human agreement on a subset if possible.
minor comments (6)
  1. [Section 3.3] There is a duplicated sentence: 'As illustrated in Figures 2b and 2c, when we cascade two medium-sized models (DeiT-B ... and TinyViT ...), each achieving around 82% accuracy individually, As illustrated in Figures 2b and 2c, ...' — remove the repetition.
  2. [Section 2.2] The phrase 'the M1 is better than that of M2' should be 'M1 is better than M2'.
  3. [Section 3.4 / Table 6] The statement that CC 'still significantly outperforms the random baseline' at severity 5 is not supported by Table 6: for fog severity 5, APGR is 0.5409, only 0.0409 above the random baseline of 0.5. Please add error bars or confidence intervals, or soften this claim.
  4. [Appendix D] The MF baseline is selected per dataset as the implementation with the highest APGR. While this is a conservative choice for the proposed method, please state explicitly in the main text which implementation (pretrained, fine-tuned, or from scratch) is used for each dataset, or report all three in the main figure.
  5. [Algorithm 2, Method 2] The pseudocode says 'top-K bins with highest average accuracy' while Section 2.3 says 'top-K confidence bins'. Reconcile this inconsistency.
  6. [General] The paper reports no error bars for key numbers and does not include a code/data release link. Please provide confidence intervals for the main quantitative claims and a reproducibility statement.

Circularity Check

0 steps flagged

No significant circularity: cascade and cleaning pipelines are fitted on validation and evaluated on held-out test splits; the load-bearing 'generalization property' is an empirical assumption, not a definitional reduction.

full rationale

We walked the derivation chain and found no step where an output is equivalent to an input by construction. Calibration parameters (temperature/Platt scaling) are fit on a validation split and evaluated on a held-out split via ECE and external benchmarks, so the calibration 'generalization' is measured, not assumed into existence. The confidence advantage a^c_i is computed on validation and used for routing; the paper's assertion 'By the generalization property of the advantage function...' is an unproved empirical generalization rather than a tautology or a fitted parameter renamed as a prediction. Its fragility is visible in the paper's own OOD results (APGR drops from 0.7984 to 0.5409 on fog severity 5), but that is a robustness/correctness concern, not circularity. The data-cleaning precision is anchored by human annotation, and the 'Cleaned ImageNet' accuracy is a benchmark-modification issue rather than a derivation-level circularity, since the cleaning and evaluation models are stated to be distinct. No load-bearing self-citation or imported uniqueness theorem appears in the cited references. Therefore the paper is not circular; it rests on an empirical generalization assumption whose limits the authors partially acknowledge through OOD evaluation.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method relies on standard calibration assumptions plus a specific unproven generalization property of the confidence advantage. No new physical entities are introduced. Free parameters are calibration and binning hyperparameters, all fit on validation.

free parameters (5)
  • temperature T (temperature scaling) = not reported
    Fitted on validation set via NLL for ImageNet calibration.
  • Platt scaling parameters a, b = not reported
    Fitted on validation set via NLL for language model confidence.
  • number of bins N = 15 (default in experiments)
    Histogram equalization bin count; sensitivity from 5 to 100 shown in appendix.
  • selection budget K (cascading) = varies; chosen on validation
    Number of top-advantage bins routed to small model; trades accuracy vs cost.
  • top-K confidence bins (cleaning) = varies
    Threshold for flagging labels; controls precision/detection trade-off.
axioms (5)
  • domain assumption Validation and test sets are drawn from the same distribution
    Standard i.i.d. assumption for supervised ML; required for calibration generalization.
  • domain assumption Confidence is monotonically related to accuracy
    Empirically observed in Fig 1, but not proven; central to calibration-based routing.
  • domain assumption Calibrated confidence is comparable across different models
    The paper assumes a given calibrated confidence level implies the same accuracy for any model; needed for confidence advantage to be meaningful.
  • ad hoc to paper The confidence advantage computed on validation generalizes to test/OOD data
    Stated as a 'generalization property' in Sec 2.2 without proof; key for routing reliability.
  • domain assumption The verifier (e.g., code execution) accurately judges correctness
    Used to define calibration targets; if verifier is noisy, calibration is biased.

pith-pipeline@v1.3.0-alltime-deepseek · 15181 in / 10703 out tokens · 94635 ms · 2026-08-03T10:56:45.497617+00:00 · methodology

0 comments
read the original abstract

When a model knows when it does not know, many possibilities emerge. The first question is how to enable a model to recognize that it does not know. A promising approach is to use confidence, computed from the model's internal signals, to reflect its ignorance. Prior work in specific domains has shown that calibration can provide reliable confidence estimates. In this work, we propose a simple, effective, and universal training-free method that applies to both vision and language models, performing model calibration, cascading, and data cleaning to better exploit a model's ability to recognize when it does not know. We first highlight two key empirical observations: higher confidence corresponds to higher accuracy within a single model, and models calibrated on the validation set remain calibrated on a held-out test set. These findings empirically establish the reliability and comparability of calibrated confidence. Building on this, we introduce two applications: (1) model cascading with calibrated advantage routing and (2) data cleaning based on model ensemble. Using the routing signal derived from the comparability of calibrated confidences, we cascade large and small models to improve efficiency with almost no compromise in accuracy, and we further cascade two models of comparable scale to achieve performance beyond either model alone. Leveraging multiple experts and their calibrated confidences, we design a simple yet effective data-cleaning method that balances precision and detection rate to identify mislabeled samples in ImageNet and Massive Multitask Language Understanding (MMLU) datasets. Our results demonstrate that enabling models to recognize when they do not know is a practical step toward more efficient, reliable, and trustworthy AI.

Figures

Figures reproduced from arXiv: 2601.07965 by Chenjie Hao, Weier Wan, Weyl Lu, Yubei Chen, Yuko Ishiwaka, Zengyi Li.

Figure 1
Figure 1. Figure 1: Reliability Diagrams Before and After Calibration. Each panel plots accuracy vs. con￾fidence across 15 equal-count bins. Light curves represent uncalibrated predictions; dark curves represent calibrated predictions. After calibration, confidence becomes more monotonic with accu￾racy and aligns more closely with the diagonal, indicating improved reliability and generalization from the calibration split to t… view at source ↗
Figure 2
Figure 2. Figure 2: Cascade Results on ImageNet-1K. Each panel shows cascade accuracy vs. the routing ratio p. Panel (a) uses a small–large model pair, while panels (b) and (c) combine models of the same size, with panel (b) using a medium–medium pair and panel (c) using a large–large pair. Cascading Results on ImageNet In the ImageNet experiments, we evaluate a single group of ImageNet-only models, consisting of a small mode… view at source ↗
Figure 3
Figure 3. Figure 3: Cascade Results on Language Models. Panels (a)–(e) show small–large model cascades on five datasets. Panel (f) shows a large–large cascade, where p denotes the fraction of samples routed to the first large model. We use LLaMA-70B as the large model and LLaMA-8B as the small model. As shown in figure 3, CC consistently outperform matrix factorization routing and the random-pick baseline across all datasets.… view at source ↗
Figure 4
Figure 4. Figure 4: OOD evaluation on ImageNet-C (Hendrycks and Dietterich, 2019) with all corruption [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: OOD Evaluation on MMLU-Adversarial As shown in the figures 4 and 5 , on vision tasks, even under the most severe corruption setting (severity = 5), the cascading approach(CC) still significantly outperforms the random baseline. On language tasks, cascading performance is likewise substantially better than the random baseline, demonstrating the robustness and generalization ability of the CC method. 3.5 DAT… view at source ↗
Figure 6
Figure 6. Figure 6: Cleaning Precision vs. Detection Rate. The horizontal axis (Cleaning Detection Rate) is defined as the ratio of flagged samples to the total dataset size. The vertical axis (Cleaning Pre￾cision) measures the fraction of truly mislabeled samples among the flagged ones, estimated using human verification. Samples are sorted by calibrated confidence in descending order and grouped into 15 equal-count bins; pr… 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

17 extracted references · 16 linked inside Pith

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Pith/arXiv arXiv

  2. [7]

    semanticscholar.org/CorpusID:234357974

    URLhttps://api. semanticscholar.org/CorpusID:234357974. Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common cor- ruptions and perturbations.arXiv preprint arXiv:1903.12261,

    Pith/arXiv arXiv 1903

  3. [9]

    Right answer, wrong score: Uncovering the inconsistencies of llm evalua- tion in multiple-choice question answering

    Francesco Maria Molfese, Luca Moroni, Luca Gioffr `e, Alessandro Scir `e, Simone Conia, and Roberto Navigli. Right answer, wrong score: Uncovering the inconsistencies of llm evalua- tion in multiple-choice question answering. InFindings of the Association for Computational Linguistics: ACL 2025,

    2025

  4. [13]

    Large language model routing with benchmark datasets.arXiv preprint arXiv:2309.15789,

    Tal Shnitzer, Anthony Ou, M´ırian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. Large language model routing with benchmark datasets.arXiv preprint arXiv:2309.15789,

    Pith/arXiv arXiv

  5. [14]

    Calibration and correctness of lan- guage models for code.arXiv preprint arXiv:2402.02047,

    Claudio Spiess, David Gros, Kunal Suresh Pai, Michael Pradel, Md Rafiqul Islam Rabin, Amin Alipour, Susmit Jha, Prem Devanbu, and Toufique Ahmed. Calibration and correctness of lan- guage models for code.arXiv preprint arXiv:2402.02047,

    Pith/arXiv arXiv

  6. [15]

    Xiaofang Wang, Dan Kondratyuk, Eric Christiansen, Kris M Kitani, Yair Alon, and Elad Eban

    URLhttps: //openreview.net/forum?id=NJS8kp15zzH. Xiaofang Wang, Dan Kondratyuk, Eric Christiansen, Kris M Kitani, Yair Alon, and Elad Eban. Wisdom of committees: An overlooked approach to faster and more accurate models.arXiv preprint arXiv:2012.01988,

    Pith/arXiv arXiv 2012

  7. [16]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    Pith/arXiv arXiv

  8. [17]

    Bigcodebench: Bench- marking code generation with diverse function calls and complex instructions.arXiv preprint arXiv:2406.15877,

    Terry Yue Zhuo, Minh Chien Vu, Jenny Chim, Han Hu, Wenhao Yu, Ratnadira Widyasari, Imam Nur Bani Yusuf, Haolan Zhan, Junda He, Indraneil Paul, et al. Bigcodebench: Bench- marking code generation with diverse function calls and complex instructions.arXiv preprint arXiv:2406.15877,

    Pith/arXiv arXiv

  9. [1999]

    Routing networks: Adaptive selection of non-linear functions for multi-task learning.arXiv preprint arXiv:1711.01239,

    Clemens Rosenbaum, Tim Klinger, and Matthew Riemer. Routing networks: Adaptive selection of non-linear functions for multi-task learning.arXiv preprint arXiv:1711.01239,

    Pith/arXiv arXiv

  10. [2017]

    Distribution-free calibration guarantees for histogram binning without sample splitting.ArXiv, abs/2105.04656,

    Chirag Gupta and Aaditya Ramdas. Distribution-free calibration guarantees for histogram binning without sample splitting.ArXiv, abs/2105.04656,

    Pith/arXiv arXiv

  11. [2018]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Pith/arXiv arXiv

  12. [2019]

    Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300,

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300,

    Pith/arXiv arXiv 2009

  13. [2021]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457,

    Pith/arXiv arXiv

  14. [2022]

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E

    URLhttps://arxiv.org/abs/1911.00068. Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. RouteLLM: Learning to route LLMs from preference data. InThe Thirteenth International Conference on Learning Representations,

    Pith/arXiv arXiv 1911

  15. [2023]

    Program synthesis with large language models.arXiv preprint arXiv:2108.07732,

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models.arXiv preprint arXiv:2108.07732,

    Pith/arXiv arXiv

  16. [2024]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Pith/arXiv arXiv

  17. [2025]

    Jeremy Nixon, Michael Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran

    13 Preprint. Jeremy Nixon, Michael Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning.arXiv preprint arXiv:1904.01685,

    Pith/arXiv arXiv 1904