cs.CLNov 16, 2026

Stability-Compensated Threshold Calibration for Selection-Conditioned Risk Control in Large Language Models

Authors: Zhiyuan Wang, Yuanchang Ye, Jiahao Yuan, Jinhao Duan, Xiaoshuang Shi, Weijun Xie, Ye Wei, Kaidi Xu

Organizations: City University of Hong Kong · National University of Singapore · East China Normal University · University of North Carolina at Chapel Hill · 

Abstract

Large language models often produce hallucinated answers, motivating selective prediction policies that abstain on uncertain queries. However, heuristic uncertainty thresholds alone do not provide statistical control of errors among accepted predictions. Linear Expectation Constraints (LEC) reduce selection-conditioned error rate (SCER) calibration to a nonpositive expected linear-loss constraint, but the associated loss is non-monotonic in the acceptance threshold. We construct an exact counterexample under i.i.d. sampling demonstrating that LEC's empirical threshold-selection rule can violate the population SCER target. To address this risk-transfer gap, we introduce Stability-Compensated Linear Expectation Control (SC-LEC), which uses the negative linear risk of a theoretical full-data reference to compensate for threshold-selection instability. For local expansion beyond a safe anchor, we derive two finite-sample certificates: a stability-based certificate exploiting rank crossings and a direct certificate using the structure of threshold expansion to reduce conservatism. With a valid population anchor margin, our certificate conditions ensure finite-sample marginal SCER control, and independent margin certification further provides a 1−δ1-δ high-probability guarantee for this calibration-averaged control. Experiments on CommonsenseQA, ARC-Challenge, and TriviaQA show useful retention gains beyond the same certified anchor, while revealing conservative behavior in small-margin regimes.

Figures & tables

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Jul 5, 2026

Uncertainty-Aware Abstention in Large Language Models with Provable Alignment Guarantees

Large language models (LLMs) are increasingly deployed in question answering (QA) systems, yet they may generate hallucinated or misaligned responses without reliable confidence estimates. Uncertainty quantification (UQ) offers a natural basis for selective answering, where a system answers only when its prediction is deemed reliable and abstains otherwise. However, existing uncertainty scores for LLMs are often heuristic: a threshold chosen on such scores does not, by itself, provide statistical guarantees on the error rate among accepted answers. We propose CIC, a confidence-interval-based calibration framework that converts arbitrary uncertainty scores into risk-controlled selective answering rules. Given a held-out calibration set, CIC evaluates each generated response using an application-specific alignment criterion and associates it with an uncertainty score and a binary error label. For each candidate uncertainty threshold, CIC estimates the acceptance-conditioned error rate and constructs a high-probability upper confidence bound using either Hoeffding-style or Clopper-Pearson confidence intervals. It then selects the largest threshold whose upper bound is below a user-specified risk level αα, thereby maximizing the answering rate subject to a finite-sample reliability constraint. Under exchangeability, CIC guarantees with probability at least 1−δ1-δ that the selected threshold, if non-null, controls the error rate among accepted answers at level αα. We evaluate CIC on both closed-ended and open-ended QA benchmarks across seven LLMs and multiple uncertainty estimators. Experimental results show that CIC consistently achieves valid risk control while retaining strong answering efficiency, providing a practical and statistically grounded mechanism for deploying LLMs in reliability-sensitive QA workflows.
Aug 12, 2026

Asymptotic Risk Calibration for Selective Question Answering

Large language models (LLMs) may generate fluent but incorrect answers, making uncertainty quantification important for reliable question answering. However, heuristic uncertainty scores cannot perfectly distinguish correct predictions from incorrect ones, and directly applying a fixed uncertainty threshold provides no statistical control over the error rate among accepted answers. To address this limitation, we propose A-CRC-QA, a post-hoc calibration framework for uncertainty-aware selective question answering. The proposed method reformulates selection-conditioned error control as a linear expectation constraint and applies a monotonized empirical-risk calibration procedure inspired by conformal risk control. Since the resulting instance-wise loss is generally non-monotone with respect to the acceptance threshold, our framework targets asymptotic rather than finite-sample risk control. A-CRC-QA is model-agnostic, requires no additional training, and can be combined with different uncertainty estimators. Experiments on CoQA and MedMCQA demonstrate its applicability to both open-ended and closed-ended question answering, achieving a favorable trade-off between accepted-answer reliability and answer retention compared with uncalibrated and confidence-bound-based baselines.
Oct 26, 2026

Reinforcement Learning Done Right for Explicit Confidence Targets

Large language models (LLMs) can produce plausible but incorrect answers when they are uncertain, which can mislead users and undermine model reliability. Training models to abstain with reinforcement learning is a natural approach, but most existing methods use a fixed error penalty and therefore cannot adapt their behavior to different confidence requirements at inference time. We study explicit confidence targets, where a user specifies a target pp in the prompt and the model directly chooses whether to answer or abstain, with the goal of achieving an accuracy of at least pp among answered questions. However, learning to respond to different confidence targets with a single model has proven difficult, and prior attempts have failed to distinguish different targets. We show that this failure comes from two problems: the reward signal that distinguishes different values of pp is often lost during optimization, while the original error penalty grows without bound as pp approaches one. We propose RLEC (Reinforcement Learning with Explicit Confidence Targets), which preserves the learning signal needed to distinguish confidence targets, introduces a stable reward family that keeps the desired answer condition unchanged, and uses reward scheduling to balance effective learning and training stability. Together, these designs enable the first successful training of a single LLM for explicit confidence target control. Extensive experiments across diverse benchmarks and model scales show that RLEC largely preserves general capabilities and achieves competitive calibration performance. RLEC also supports fine-grained target control without an external decision system or threshold calibration after training. By making explicit confidence targets part of the trained policy, RLEC opens a new and more direct path toward behavioral calibration.