Researchers Develop Introspection Adapters to Make LLMs Report Their Learned Behaviors

A new research paper introduces "introspection adapters" (IAs) as a scalable method for making large language models (LLMs) describe their learned behaviors in natural language. The approach addresses the challenge of detecting unexpected, harmful, or hard-to-identify behaviors that can emerge when models are fine-tuned. Researchers create training data by fine-tuning models from a shared base LLM with implanted behaviors, then train a single LoRA adapter across these fine-tuned models to enable them to verbalize those behaviors. This IA technique generalizes effectively, even to models trained in very different ways from the training set, and achieves state-of-the-art performance on AuditBench for identifying explicitly hidden concerning behaviors. The method aims to simplify auditing processes for model developers and users by allowing LLMs to self-report their internal behaviors. The research was published on arXiv under the identifier arXiv:2604.16812v1.