Confidence Estimation in Language Models

Overview

The field studies whether language models can tell when their answers are likely to be wrong, which matters for deciding when to answer, abstain, retrieve evidence, or spend more computation. Methods include verbalized confidence, token probabilities and sampling-based agreement, as well as hidden-state probes, activation signals, post-hoc calibration, and conformal prediction. These approaches are tested across settings such as factual question answering, mathematical reasoning, code generation, multilingual tasks, and tool-using agents. A central challenge is that confidence signals can disagree or shift with prompts, training, and distribution changes; good calibration or error ranking alone does not guarantee useful decisions or reliable risk guarantees.

Latest papers

271 papers39 papers in the last four weeks, down 13% on the four weeks before. 0.3% of all new papers.

May 23, 2025

Popular but Wrong: Understanding and Mitigating LLM Overconfidence through Knowledge Popularity

Large language models (LLMs) often produce incorrect answers with high confidence, yet the factors associated with such overconfidence remain insufficiently understood. We study this problem through the lens of knowledge popularity. Using entity-centric factual QA derived from Wikidata triplets, we characterize popularity through question entity popularity, answer popularity, and question-answer co-occurrence. We find two consistent patterns. First, hallucinated answers are far from random: compared with ground-truth answers, they tend to be more popular or more frequently associated with the question entity. Second, confidence is strongly tied to the popularity of generated answers: even among incorrect predictions, more popular answers or those with higher question-answer co-occurrence receive higher confidence. Together, these findings suggest that popular but wrong alternatives may contribute to overconfidence. We further show that popularity-related signals can mitigate overconfidence and improve overall confidence estimation. Across six models and three datasets, incorporating knowledge popularity reduces average confidence on incorrect answers from 0.765 to 0.254 and overall ECE from 0.356 to 0.050, while improving Alignment from 77.08% to 83.72%.