EMNLP 2026 Main Conference | Peking University × Yixin AI Lab: Why Do LLMs Answer Wrongly When They Already 'Know'?

A collaboration between Peking University and Yixin AI Lab found that when large language models answer logic questions incorrectly, it is often not due to a lack of reasoning ability. Instead, the correct answer is already encoded in hidden states but is masked by systematic bias in the output layer when converting to candidate scores, a phenomenon the authors call the 'Readout Bottleneck.' The study proposes a staged diagnostic method; on Qwen3.5 models, probe accuracy reaches 0.830 while native sequence scoring drops to 0.333, with 999 out of 1,000 questions assigned the same label. The paper has been accepted to the EMNLP 2026 main conference.

+5