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.
At the Silicon Photonics Development Forum of the 19th Pujiang Innovation Forum, Tao Yaoyu from Peking University's Xinyao Yue team unveiled a programmable spatial optoelectronic heterogeneous AI inference chip. It adopts a digital light intensity accumulation and multiply-accumulate split short optical path architecture, targeting low-error, low-loss and programmable AI inference acceleration, achieving first-token latency at the hundred-microsecond level.