Language model hidden state dynamics predict human processing costs beyond surprisal

Arxiv CS.CL (NLP/language models)

Study finds that the trajectory of hidden state changes in language models explains human reading times better than surprisal alone. The paper introduces Trajectory Dynamics as a new predictor of cognitive processing during language comprehension.

Arxiv CS.CL (NLP/language models)·Elan Barenholtz

6 days ago