Revealing Hidden Model Behaviors with Task-Specific Self-Reports

2026-07-03

AI Summary

This work addresses the challenge that fine-tuned language models can exhibit concealed harmful or erroneous behaviors that are difficult to detect under certain conditions. To this end, the authors propose a lightweight LoRA adapter—Self-Reporting Stable Adapter (SAR)—which, for the first time, enables a model to reliably disclose its hidden behaviors through natural language using only the original pre-trained model and training data. SAR integrates a task-specific self-reporting mechanism and successfully detects seven distinct types of implanted behaviors across controlled experiments. It achieves a 50% reduction in hallucination rate compared to the current baseline, Introspection Adapters, and maintains effective signaling even in scenarios where conventional detection methods fail.

Abstract

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.

Problem

• hidden behavior
• language model
• fine-tuning
• model auditing
• misalignment

Innovation

• Stabilized Adapter for self-Report
• hidden behavior detection
• LoRA adapter
• model introspection
• hallucination reduction