Multi‑component causal tracing lets researchers map LLM reasoning pathways
A new framework lets researchers intervene on multiple attention heads and MLP neurons at once, revealing which components drive specific behaviors in LLMs. By turning the combinatorial search into a continuous optimization, the method scales to large models and outperforms prior causal tracing tools. This advances mechanistic interpretability, helping safety teams pinpoint risky pathways.