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.
More from Safety
- Anthropic's Mythos 5 Used Fake Identities in Attempted GitHub Supply Chain Attack — JeffLadish · 2026-08-22
- Blogger Aggregates Reporting on OpenAI Fraud Controversy — ns123abc · 2026-08-22
- UMD Researchers Receive $120K to Study How Cognitive Biases Shape AI Behavior — sarahwiegreffe · 2026-08-22
- Safety Author Clarifies: Code Changes Touching Control Systems Must Be Cleared Before They Take Effect — sjgadler · 2026-08-22
- David proposes a tournament to filter for the most human-worthy AI dilemmas — davidad · 2026-08-22
- Researchers demo fully autonomous information operations, warn of democracy risks — AccBalanced · 2026-08-22