Machine Heart

Sep 18, 2026 · Artificial Intelligence

Readout Bottleneck: LLMs Know Answers But Fail to Output Them

A Peking University and Yixin AI Lab paper accepted at EMNLP 2026 reveals that large language models often encode correct reasoning in hidden states but suffer a 'readout bottleneck' where output-layer biases collapse final predictions; a two-parameter calibration recovers up to 30+ accuracy points without retraining.