Circuits, Gradients, and Tool Trust
Published:
We tried to scale mechanistic interpretability to a real safety task — one where a model must decide whether to trust a wrong tool response. Attribution graphs, run through a circuit oracle, performed at 67% accuracy. A single backward pass at the correct-tool prompt reached 85%.
This post is published on LessWrong. Joint work with Sriram Balasubramanian (SPAR SP26).