AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-19
Deep LLM layers expose passage-level context in EEG during silent reading
A previous study decoded word-level information from 19-channel EEG recorded during silent reading using LLM hidden-state embeddings. This study trained one decoder per layer of Llama-3.1-8B and found that word-level decodability peaks in early layers and declines toward the last layer, while context tracking rises with depth. Controls indicated that the context tracking mostly follows the text rather than word decoding or session-specific signatures, and even a single 0.5-second EEG window carried passage-level information.