Peking University and YIXIN AI Lab Paper on LLM 'Readout Bottleneck' Accepted at EMNLP 2026

A joint research paper by Peking University and YIXIN AI Lab, titled "Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores," has been accepted to the main conference of EMNLP 2026, a top venue for natural language processing. The acceptance rate is 15.4%.

The study introduces a diagnostic perspective called the "Readout Bottleneck" for evaluating large language model reasoning. It asks whether a model that fails a logic question truly failed to reason, or whether the correct answer was encoded internally but masked by systematic biases during output generation.

The researchers found that linear probes can decode the correct answer from the model's hidden states with high accuracy, both at the end of the prompt and immediately after the "Answer:" token. However, when the information passes through the vocabulary projection and is aggregated into full sequence scores, accuracy drops sharply. For example, on the lexical-OOD slice, Qwen3.5-9B achieved a probe accuracy of 0.830, while its native sequence scoring accuracy fell to 0.333, close to random guessing in a three-class setting. Across 1,000 test items, native sequence scoring assigned 999 items to the same label, suggesting a strong global bias in the output layer.

The paper also proposes a minimal correction method that fits only two parameters via grid search on unlabeled context scores, aligning the predicted distribution with class priors. On synthetic logic tasks, Qwen3.5-4B and Qwen3.5-9B accuracy improved from 33.3% to 57.0% and 60.2%, respectively. On the ProofWriter natural language reasoning task, accuracy rose from 33.3% to 65.3% and 67.8%. The authors also used TF-IDF-missed hard examples and a count-preserving permutation null to rule out shallow lexical shortcuts and random fluctuations.

The work offers a new perspective on how LLM benchmark results should be interpreted, suggesting that near-random scores do not necessarily mean a lack of reasoning ability, and that hidden states should be examined before concluding that a model "cannot do" a task.