It is found that signal effectiveness is task-dependent: confidence is strongest on MCQ and simpler math, while likelihood/KL signals give the most frequent gains on harder math and code; no signal is universally best across model updates either, and some cross-version signals stay informative even when confidence fails, including without labels.
Abstract
Frontier LLMs are updated frequently and typically outperform their predecessors in aggregate. But aggregate gains say little about individual samples: an update can still cause sample-level regression, where a response correct under the old model becomes incorrect under the new one. This paper studies how to predict such regressions from signals available at inference time. We compare single-model signals (confidence, logit margin, attention entropy) against cross-version signals (output KL divergence, likelihood drift, token-level KL, representation drift) under a unified added-value test that isolates each signal's gain over a confidence baseline. Across six benchmarks in three task families (multiple-choice question answering, or MCQ; math reasoning; code generation) and six model update pairs, we find that (1) signal effectiveness is task-dependent: confidence is strongest on MCQ and simpler math, while likelihood/KL signals give the most frequent gains on harder math and code; (2) no signal is universally best across model updates either; and (3) some cross-version signals stay informative even when confidence fails, including without labels, which supports a proof-of-concept selective fallback that routes high-risk samples back to the old model. Practitioners can use these task-level patterns to choose which regression signal to trust for a given update. Code is available at https://github.com/jiashengsally/llm-regression-signals.
It is proposed that a model's uncertainty about a token is reflected not only in the breadth of its output distribution but also in whether a confident prediction is \emph{fragile} under perturbation of its attention pathways, a training-free estimator that masks attention heads and measures the BALD mutual information among the resulting subnetworks.
Minsoo Kim, Sungyoung Ji, Kisung Moon et al.· 0 citations
Aggregate accuracy hides where models succeed and fail. Estimating conditional performance profiles from gold labels alone is expensive, while cheap auxiliary signals such as LLM-judge scores, pairwise comparisons, confidence scores, and judge-disagreement features can be collected for every benchmark item but are often biased or miscalibrated. We propose LACE (Local Augmented Control-Variate Evaluation), a semi-supervised estimator for conditional LLM evaluation. The key step is local centering: after subtracting the conditional mean of a cheap signal within the target profile region, any linear augmentation has zero conditional mean and therefore cannot change the estimand. The augmentation coefficient is used only for efficiency, and a local ridge control variate combines a gold-label residual mean from the labeled subset with a cheap-signal mean from the full item pool. We prove calibration-free identification, unbiasedness for grouped profiles, local oracle optimality within centered linear augmentations, and first-order adaptivity to the estimated coefficient. The resulting gain formula is governed by a population local $R^2$, which characterizes how the efficiency attainable from the cheap signals varies across profile values. We also derive corresponding estimators for direct paired model gaps and deployment-weighted scores. We empirically evaluate the primary performance-profile estimator on MATH-500, ScienceQA, MMLU, WinoGrande, HellaSwag, TruthfulQA, GSM8K, and ARC.
Zhi Zhang, Lingfeng Lyu, Yue Kang et al.· 0 citations
Item Response Theory (IRT) has recently been proposed as a framework for evaluating large language model (LLM) benchmarks by separating a model's latent ability from the properties of individual benchmark items. Existing neural IRT approaches, including PSN-IRT, estimate these quantities using point estimates, limiting uncertainty quantification and downstream statistical inference. We introduce Laplace-PSN-IRT, a post-hoc last-layer Laplace approximation that augments a trained PSN-IRT model with approximate Bayesian posterior inference, recovering calibrated uncertainty over model ability and item difficulty without retraining. The resulting posterior enables credible intervals, probabilistic comparisons between models, and propagation of parameter uncertainty into Fisher-information-based item selection. We show that most pairwise comparisons among 12 models on a standard LLM benchmark leaderboard are not statistically distinguishable despite differing point-estimate ranks. We further show that point-estimate Fisher information can become nearly zero for many benchmark items because it is evaluated at a single reference ability, whereas posterior-expected Fisher information remains substantially more stable across the ability range. Finally, posterior-expected Fisher information more accurately recovers full-benchmark ability rankings from small benchmark subsets in most experimental settings while matching point-estimate performance for the smallest subsets. We validate the calibration of the approximate posterior using held-out predictive coverage and find that modeling item difficulty as random while treating item discrimination as fixed produces well-calibrated uncertainty in this architecture.
Juan Francisco, Mandujano Reyes· arXiv.org· 0 citations
This work proposes CORA-Diff, a training-free method that preserves the original transfer rule and applies confidence-and-persistence gating only to positions that rule leaves unresolved, and shows that native confidence and persistence enable reliable residual acceptance, reducing repeated denoising computation while preserving task quality.
Large language models have made strong reasoning gains through supervised fine-tuning, reinforcement learning, and on-policy distillation, yet these post-training methods are usually evaluated only by final-answer accuracy. We study how they reshape confidence during reasoning. We introduce a three-stage calibration framework that evaluates confidence before, during, and after chain-of-thought generation, corresponding to difficulty estimation, early termination, and answer aggregation. Through a controlled comparison on mathematical reasoning benchmarks, we find that OPD provides the most useful pre-reasoning confidence, SFT gives the strongest online signal for early stopping, and RL produces the most reliable trace-level signal for aggregation. We further show that confidence reliability is position-dependent: RL confidence becomes informative after a path-commitment phase, while OPD confidence is useful early but can become inversely calibrated later. Based on this observation, we propose PosConf, a position-aware confidence strategy that uses confidence only from reliable relative-position intervals. PosConf improves RL answer aggregation by 6.1 points over majority voting and consistently improves OPD early stopping under tight token budgets, with gains up to 4.3 points by avoiding its later inverse-calibration region, showing that \emph{confidence in reasoning models should be used both stage-wise and position-awarely}. Our code is available at https://github.com/EIT-NLP/Post-Training-Calibration.
Shuhao Li, Guodong Du, Anhao Zhao et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
