Yohei Oseki

Fine-tuned neural language models can accurately predict human reading times for garden-path sentences, challenging the notion that surprisal cannot account for syntactic disambiguation.

Later layers of LLMs capture cognitive effort in syntactically challenging sentences better than earlier layers, but still miss the mark compared to human processing.

Forget everything bad: a new unlearning method wipes away harmful LLM behaviors by selectively preserving only the knowledge you want.