Large language models are expected to estimate their own uncertainty, with confidence signals used to guide retrieval, adaptive compute allocation, routing, and selective abstention. We investigate wh...
arXiv:2607. 07626v1 Announce Type: cross Abstract: Reliable confidence estimation is essential for deploying large language models (LLMs) in confidence-aware systems, where downstream decisions such as retrieval, tool use, and adaptive computation depend on accurately estimating answer reliability.
arXiv:2603.17839v4 Announce Type: replace-cross Abstract: Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is widely used to extract uncertainty estimates from...
The study trains ten open‑weight large language models (LLMs) to predict their own accuracy on factual multiple‑choice questions before answering. Results show that the models’ confidence signals split into two distinct patterns: early in training, confidence aligns with output consistency (how concentrated the answer distribution is), while later, it aligns with true accuracy but only on data similar to the training set. This indicates that calibration training may not universally teach LLMs to detect their own errors.
arXiv:2604.08974v2 Announce Type: replace Abstract: Uncertainty quantification techniques measure confidence in language model outputs to support critical applications like hallucination detection an...
arXiv:2508. 14390v2 Announce Type: replace-cross Abstract: Large language models (LLMs) often express verbal confidence that is poorly aligned with actual correctness, limiting their reliability in safety-critical applications.
arXiv:2607. 20526v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed in settings where fluent but incorrect answers can be costly.
The paper introduces a framework for assessing how well language models (LMs) calibrate their confidence when answering relational knowledge questions. It evaluates three types of confidence—intrinsic, paraphrase consistency, and linguistic—across 21 causal and 8 masked LMs, finding that most models, especially masked ones, are overconfident. The study shows that confidence estimates that account for rephrasing inconsistencies and increasing answer options lead to better calibration.
The paper investigates whether language models’ self-reported confidence is meaningful without additional training. By evaluating three training‑free signals—direct verbalization, post‑hoc probability estimates, and agreement across multiple generations—on 100 TriviaQA questions, the authors find that direct verbalization alone achieves high AUROC scores (0.956 and 0.937) for correctness prediction, while agreement-based methods perform noticeably worse. Re‑eliciting confidence for the same answers shows modest score shifts and occasional decision flips, and an audit of biography claims reveals only a small confidence gap between supported and contradicted statements.
arXiv:2608.28382v1 Announce Type: new Abstract: Users often ask large language models (LLMs) to report how confident they are, but it is unclear whether such linguistic confidence tracks the model's...
The paper introduces TASCO, a test‑time adaptation framework that enhances Large Language Model reasoning by incorporating local stability into confidence‑based adaptation while keeping the model frozen. TASCO optimizes a lightweight task‑level prefix using two perturbation strategies—Random Perturbation for distributional stability and Sharpness‑Aware Perturbation for worst‑case sensitivity—to ensure that high confidence aligns with correctness. Experiments show that TASCO improves reasoning accuracy and token efficiency across various LLMs and benchmarks, and behavioral analyses confirm that it maintains stable confidence without over‑concentrating the predictive distribution.
PROOF is a benchmark that profiles the reliability of object-level facts in instruction-tuned language models by converting a frozen Wikidata snapshot into 18,486 English multiple-choice questions grounded in 11,779 semantic facts across 101 classes, 392 properties, and 14 domains. Each question includes an explicit "I don't know" option, a "No correct option" control, and nine controlled formulations, with 1,849 questions designed as no-correct-option traps. The study evaluates 18 open-weight model deployments on 166,374 prompts, revealing wide variability in factual accuracy, sensitivity to wording changes, and the impact of decoder perturbations.