Anthropic evaluates LLM behavior explanations via counterfactual experiments

saprmarks · x · 2026-08-22

Anthropic published a research blog introducing CHIVE, an agentic pipeline that discovers unexpected LLM behaviors in the wild and explains them via counterfactual prompt edits. The study found that activation-reading interpretability tools provided no uplift in predicting experiment outcomes compared to just reading the transcript. However, models trained on this data to predict behavior changes from prompt edits showed generalization to held-out settings.

Related event: Anthropic study: current interpretability tools fail to beat just reading transcripts(14 posts)→

Original post →

More from Safety

Safety channel →