← Back to all news
arXiv Computation and Language By Nicolas Yax, Stefano Palminteri, Pierre-Yves Oudeyer

LLMs learn different forms of metacognition when trained to predict their own accuracy

Read the original on arXiv Computation and Language →

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.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computation and Language.

Related stories

arXiv AI
1d ago

Certainty Is Not Just Correctness: Rethinking Token-Level Certainty in LLM Reasoning

The paper investigates token‑level certainty as a proxy for correctness in large language models. It finds that certainty better predicts whether a model will answer a question correctly than it does whether a specific response is correct, and that certainty varies by token type and position. The authors show that using certainty early in generation to allocate responses and later to weight votes improves accuracy while dramatically cutting token cost.

By Yunfan Zhou, Ye Zhu, Zhihai Wang, Jianguo Yao, Haibing Guan, Xijun Li
llms
More like this →
arXiv AI
Sep 25

PROOF: Profiling Reliability of Object-Level Facts in Large Language Models

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.

By Andrei Chetvergov, Mikhail Solovev, Timofei Sivoraksha, Stepan Ukolov, Valeriia Kuschenko, Alexander Evseev, Sergey Bolovtsov
llmsbenchmarkssafety
More like this →
arXiv AI
Aug 17

The Metacognitive Bottleneck: Japanese Riddles Reveal Fundamental Limits of Machine Insight and Self-Evaluation in Reasoning AI

arXiv:2509. 14704v3 Announce Type: replace Abstract: Benchmark saturation and training-data contamination increasingly obscure whether reported gains in large language models (LLMs) reflect genuine advances in reasoning or familiarity with recurring patterns in benchmark problems.

By Masaharu Mizumoto, Dat Nguyen, Zhiheng Han, Xingfu Li, Yo Nakawake, Le Minh Nguyen
llmsbenchmarks
More like this →
arXiv Computation and Language
Sep 18

An Analysis of Training-Free Self-Reported Confidence in Language Models

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.

By Lukas Meyer, Sofia Rossi, Wei Chen, Thomas Laurent, Yiming Li
llmsbenchmarks
More like this →