Ameisen et al.

New training method enhances LLM interpretability by reducing signal loss

Researchers have developed a new method called replacement-aware training to improve the interpretability of large language models. This technique trains sparse auto-encoders (SAEs) to be robust to errors introduced by …