The Calibration Floor: Format Repair Can Masquerade as Self-Correction at Small-to-Mid Scale

• URL: http://arxiv.org/abs/2608.04355v1
• Date: Wed, 05 Aug 2026 01:55:33 GMT
• Title: The Calibration Floor: Format Repair Can Masquerade as Self-Correction at Small-to-Mid Scale
• Abstract summary: Accuracy changes after language-model self-revision are usually interpreted as changes in reasoning.<n>We show this can fail at the answer-extraction boundary, and test the failure causally rather than only observationally.
• Score: 7.159691098846351
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Accuracy changes after language-model self-revision are usually interpreted as changes in reasoning. We show this can fail at the answer-extraction boundary, and test the failure causally rather than only observationally. Across Qwen3.5 (0.8B-9B), Gemma-4-12B, and two frontier models via API (Tencent Hy3, Nvidia Nemotron-3-Ultra-550B) in 29 primary cells plus a frontier arm, we decompose the always-revise accuracy shift into a content margin (both answers parseable) and format-recovery/loss margins (parseability changes). On 12 cells with meaningful unparseable-answer rates, format effects exceed content effects (Wilcoxon p=1.7e-3). To test this causally, we force already-generated reasoning through grammar-constrained decoding so every answer is parseable by construction: across 14 cells this closes a median 71% of the gap between the naive total effect and the content-margin estimate, with two cells converging exactly and a residual on the two largest-effect cells reported rather than dismissed. A clustered model confirms floor-scale (0.8B/2B) models have far higher odds of content-level change and harm than capable-scale models (p<1e-7). Replicating a cited confidence-gating protocol verbatim on Qwen3.5 does not reproduce its reported gain and shows the same near-zero content margin. A frontier check on much larger models shows format-dominance intensifying with scale: content margin is exactly zero in all 5 cells despite total effects up to +0.275, though this arm is lower-powered. The calibration-floor criterion on the content margin reveals a squeeze: floor-scale cells have headroom but insufficient signal, capable-scale cells have signal but little headroom; only one cell is marginally viable, with negligible sealed-holdout gain. Content is a minority share of what the field has measured as self-correction. We release the instrument, code, and derived results.

Related papers

• Cross-Block Conditioning in Deep Boltzmann Machines for Statistical Data Fusion [0.0]
  We propose observed-block multi-prediction, which restricts the multi-prediction objective to targets drawn from what each row actually observes.<n>Having a discriminative criterion that survives the setting lets us ask whether the joint model is needed at all.
  arXiv  Detail & Related papers  (2026-09-14T02:26:16Z)
• When Do Internal Probes Beat Reading the Answer? Miscalibrated Readouts and Behavior-Concealed Knowledge in Language Models [0.0]
  A language model, asked to verify 1,200 logical conclusions, answers YES every time.<n>linear probes on its hidden states read the correct verdict at 0.96 AUC.<n>We ask where the verdict is lost, and find the dominant failure is a single scalar.<n>The diagnosis generalizes: across 90 semantic-label configurations of a five-model, three-family factorial, behavioral accuracy collapses onto a single function of threshold offset.
  arXiv  Detail & Related papers  (2026-09-04T00:26:45Z)
• A Calibration Audit of Confidence in Feed-Forward 3D Reconstruction [45.496856835431885]
  3D reconstruction models emit per-pixel confidence that downstream systems read as a reliability signal.<n>We score the confidence on four properties, how well it ranks error, whether its level is right on average, whether it holds across the confidence range, and whether its intervals cover the truth.<n>The confidence ranks error well, but the predicted uncertainty is too low when it is read under conditions that are not exactly those of training.
  arXiv  Detail & Related papers  (2026-08-30T10:21:34Z)
• Feature Priming in Online Linear Regression: Sparse-Regret Lower Bounds and a Tight Univariate Rate [47.36630149538994]
  In high-dimensional online prediction, the best predictor may depend on only a few features, so regret should scale with sparsity rather than the ambient dimension.<n>Warmuth and Amid asked at COLT 2023 whether any of three such rules admits a competitive online regret guarantee.<n>Our analysis identifies a common obstruction: cheap nuisance causes the refit to underweight the truly predictive coordinate.
  arXiv  Detail & Related papers  (2026-08-18T09:32:03Z)
• Safety-Gated Agentic Supervisory Control on a Coupled Distillation Benchmark: Regime Map, Auditable Gate, and Co-Design Findings [0.0]
  This paper puts that check in a rule-based forked-twin counterfactual gate.<n>On Skogestad's Column A the ladder is PID-only (C0), linear MPC (C1), ungated agent (C2), and gated agent (C3)
  arXiv  Detail & Related papers  (2026-07-30T08:26:53Z)
• Confidence Scores in Open-Vocabulary Detection Are a Biased Mixture of Scale and Semantics [8.465978346858842]
  Foundation models such as CLIP have enabled open-vocabulary object detectors that generalise to novel categories via vision-language similarity.<n>We show that the scale-bias finding is a biased mixture of two effects.<n>These findings reveal a fundamental limitation of adapting image-level foundation models to region-level detection tasks.
  arXiv  Detail & Related papers  (2026-07-13T01:33:18Z)
• Score Distributions, Not Cells: Evaluating Single-Cell Perturbations Under Class Overlap [0.09558392439655013]
  Two perturbations can produce different populations of cells while overlapping so much that an individual cell could belong to either.<n>Per-cell accuracy measures this overlap rather than model quality.<n>The fix is to score perturbations across the whole population rather than cell by cell.
  arXiv  Detail & Related papers  (2026-07-06T01:47:02Z)
• CAVEWOMAN: How Large Language Models Behave Under Linguistic Input and Output Compression [77.41779716950387]
  We present Cavewoman, a two-channel evaluation protocol that scores every generation on task accuracy, realized per-item cost, and reference-text agreement.<n>We evaluate eight models on five datasets at five reduction levels, with both channels measured on the same items.
  arXiv  Detail & Related papers  (2026-06-23T02:56:48Z)
• Output Vector Editing for Memorization Mitigation in Large Language Models [68.30351930772788]
  Large language models memorize and reproduce sequences from their training data, creating privacy, copyright, and security risks.<n>Existing neuron-level mitigation methods equate editing with zeroing out neuron activations, but the activation only controls whether a neuron engages.<n>We propose output vector editing, a constrained-optimization edit that locates a small set of neurons responsible for a memorized continuation.
  arXiv  Detail & Related papers  (2026-06-17T07:29:18Z)
• BibTeX Citation Hallucinations in Scientific Publishing Agents: Evaluation and Mitigation [34.429649156970015]
  Prior evaluations tested base models without search, which does not reflect current practice.<n>Three search-enabled frontier models generate Bib entries scored on nine fields and a six-way error taxonomy.<n>Overall accuracy is 83.6%, but only 50.9% of entries are fully correct.
  arXiv  Detail & Related papers  (2026-04-03T16:30:58Z)
• PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling [32.464109652332915]
  We introduce PerturbDiff, which shifts modeling from individual cells to entire distributions.<n>PerturbDiff achieves state-of-the-art performance in single-cell response prediction and generalizes substantially better to unseen perturbations.
  arXiv  Detail & Related papers  (2026-02-23T10:28:56Z)
• On the Entropy Calibration of Language Models [52.47557449370603]
  We study the problem of entropy calibration, which asks whether a language model's entropy over generations matches its log loss on human text.<n>We find that the observed scaling behavior is similar to what is predicted by the simplified setting.<n>We prove that it is possible, if we assume access to a black box which can fit models to predict the future entropy of text.
  arXiv  Detail & Related papers  (2025-11-15T00:33:03Z)
• MeCaMIL: Causality-Aware Multiple Instance Learning for Fair and Interpretable Whole Slide Image Diagnosis [40.3028468133626]
  Multiple instance learning (MIL) has emerged as the dominant paradigm for whole slide image (WSI) analysis in computational pathology.<n>textbfMeCaMIL, a causality-aware MIL framework, explicitly models demographic confounders through structured causal graphs.<n>MeCaMIL achieves superior fairness -- demographic disparity variance drops by over 65% relative reduction on average across attributes.
  arXiv  Detail & Related papers  (2025-11-14T06:47:21Z)
• HyFormer-Net: A Synergistic CNN-Transformer with Interpretable Multi-Scale Fusion for Breast Lesion Segmentation and Classification in Ultrasound Images [0.0]
  HyFormer-Net is a hybrid CNN-Transformer for simultaneous segmentation and classification with intrinsic interpretability.<n>On the BUSI dataset, HyFormer-Net achieves Dice Score 0.761 +/- 0.072 and accuracy 93.2%, outperforming U-Net, Attention U-Net, and TransUNet.<n>We conduct the first cross-dataset generalization study for hybrid CNN-Transformers in breast ultrasound.
  arXiv  Detail & Related papers  (2025-11-02T17:06:12Z)
• Confidence-Based Model Selection: When to Take Shortcuts for Subpopulation Shifts [119.22672589020394]
  We propose COnfidence-baSed MOdel Selection (CosMoS), where model confidence can effectively guide model selection. We evaluate CosMoS on four datasets with spurious correlations, each with multiple test sets with varying levels of data distribution shift.
  arXiv  Detail & Related papers  (2023-06-19T18:48:15Z)
• Semi-Supervised Learning with Pseudo-Negative Labels for Image Classification [14.100569951592417]
  We propose a mutual learning framework based on pseudo-negative labels. By reducing the prediction probability on pseudo-negative labels, the dual model can improve its prediction ability. Our framework achieves state-of-the-art results on several main benchmarks.
  arXiv  Detail & Related papers  (2023-01-10T14:15:17Z)
• Training image classifiers using Semi-Weak Label Data [26.04162590798731]
  In Multiple Instance learning (MIL), weak labels are provided at the bag level with only presence/absence information known. This paper introduces a novel semi-weak label learning paradigm as a middle ground to mitigate the problem. We propose a two-stage framework to address the problem of learning from semi-weak labels.
  arXiv  Detail & Related papers  (2021-03-19T03:06:07Z)
• TACRED Revisited: A Thorough Evaluation of the TACRED Relation Extraction Task [80.38130122127882]
  TACRED is one of the largest, most widely used crowdsourced datasets in Relation Extraction (RE) In this paper, we investigate the questions: Have we reached a performance ceiling or is there still room for improvement? We find that label errors account for 8% absolute F1 test error, and that more than 50% of the examples need to be relabeled.
  arXiv  Detail & Related papers  (2020-04-30T15:07:37Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.