New LLM-Driven Feature Discovery method extracts 20k behavior features from chat transcripts
No model internals needed—just one LLM call per prompt yields unsupervised feature discovery.
Researchers Josh Engels, bilalchughtai, and Neel Nanda introduce LLM-Driven Feature Discovery, a method that uses a black-box LLM autorater to extract behavioral features from chat transcripts. Analyzing 100k Gemini transcripts, they generated 20k user, thought, and response features—20k features total across those categories. Features are clustered and labeled by an LLM, revealing interesting Gemini behaviors. The approach is simpler than existing methods like Explaining Datasets in Words, requiring no iterative optimization or supervised targets.
- Analyzed 100k Gemini chat transcripts, generating 20k features each for user turns, model thoughts, and assistant responses.
- Uses a single LLM autorater call per prompt with no need for model internals or iterative optimization.
- Clusters revealed interesting behaviors like sycophancy, hallucination, and roleplaying, but couldn't predict specific thoughts or responses from user features alone.
Why It Matters
Opens a new way to audit AI behavior without accessing model internals, valuable for alignment and safety research.