Findings of the BlackboxNLP 2025 Shared Task: Localizing Circuits and Causal Variables in Language Models

Dana Arad, Yonatan Belinkov, Hanjie Chen +5

Mechanistic interpretability (MI) seeks to uncover how language models (LMs) implement specific behaviors, yet measuring progress in MI remains challenging. The recently released M…