Revealing Hidden Model Behaviors with Task-Specific Self-Reports
For practitioners auditing fine-tuned models, SAR provides a reliable method to uncover hidden misalignment, reducing false reports compared to existing introspection methods.
The paper introduces SAR, a lightweight LoRA adapter that enables fine-tuned language models to describe their own hidden behaviors in plain language. Across seven implanted behaviors, SAR detects all of them, outperforming the baseline IA which misses some and hallucinates others, halving the hallucination rate.
Fine-tuning can give a language model a hidden behavior--it may give false answers under a narrow condition, or give harmful advice only when a prompt touches a particular topic. We introduce the Stabilized Adapter for self-Report (SAR), a lightweight LoRA adapter that makes a fine-tuned model describe its own hidden behavior in plain language, using only the model and the dataset it was trained on. Across seven implanted behaviors (plus a no-behavior control), SAR detects the hidden behavior in every one--even when the model has generalized into broad misalignment that the training data alone does not predict. Introspection Adapters (IA), the closest existing baseline, detects some behaviors from our suite but misses others entirely--and where it misses, it hallucinates, consistently reporting wrong behaviors. SAR retains positive signal on every setting where IA fails and halves the rate of hallucinations. This makes it much easier for practitioners to audit their models and obtain reliable answers to "what did my model actually learn?" type of questions.