Anthropic’s CHIVE pipeline tests whether interpretability tools actually help explain model behavior
Anthropic’s Alignment Science team introduced CHIVE, an agentic pipeline that searches for unexpected model behaviors in real transcripts and then probes them with counterfactual prompt edits. In the company’s August 21 research post, Anthropic says CHIVE produces two outputs for each case: a natural-language explanation and the measured results of the prompt-edit experiments. The key reported finding is that activation-reading interpretability tools did not improve performance on this evaluation; agents with those tools predicted the outcomes no better than agents that only saw the transcript. Anthropic also says the CHIVE data can be used as training data, and that models trained to predict whether prompt edits would change behavior generalized to held-out settings. The work is presented as Anthropic’s research result, not an independent benchmark.
AI-generated summary with automated source checks. It has not been reviewed by a human; consult the linked evidence.
Read original at The Sequence ↗
Originally published 10/22/2026, 11:11:37 AM