A research paper by Peking University and YIXIN AI Lab, 'Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores,' was accepted to the main conference of EMNLP 2026 (acceptance rate 15.4%). The study proposes a 'readout bottleneck' diagnostic perspective for LLM reasoning evaluation, showing that models may encode correct answers in hidden states while systematic output-layer biases obscure them, and designs a staged diagnostic method to verify the phenomenon.