New technique helps AI language models become more honest about their confidence levels without needing labeled training data

Large language models are overconfident and frequently express high certainty on questions they answer incorrectly

Researchers discovered LLMs already contain better-calibrated signals internally—their probability of answering 'True' to 'Is this answer correct?' outperforms their stated confidence

SECL (Self-Calibrating Language Models) uses test-time training to exploit this gap as label-free self-supervision, requiring no labeled data or human intervention

The method adapts only when input distribution shifts occur, avoiding the high inference costs of existing calibration approaches