Trajectory Dynamics in Language Model Hidden States Predict Human Processing Costs Beyond Surprisal
Researchers have introduced a new metric called trajectory extrapolation error to better predict human language processing costs. This metric analyzes the trajectory of hidden states in transformer language models, going beyond traditional surprisal measures. The findings indicate that this new metric independently predicts reading times and is particularly effective for complex sentence structures, strengthening with larger model scales. AI
IMPACT Introduces a novel metric for understanding language model behavior and its relation to human cognition.