Introspection Adapters: Training LLMs to Report Their Learned Behaviors

Keshav Shenoy, Li Yang, Abhay Sheshadri, Sören Mindermann, Jack Lindsey, Sam Marks, Rowan Wang · 2026 · arXiv preprint

Read the source

A single adapter trained to make fine-tuned models state their implanted behaviors generalized to unseen models and set the best score on AuditBench.

Questions it bears on

• An agenda beyond the Assistant
• How well can LLMs introspect