Introspection Fine-Tuning (IFT): Training Small LLMs to Introspect

Small language models can be trained to detect their own internal changes: The research demonstrates that models as small as 2 billion parameters can learn to identify when their internal processing has been artificially altered, and this ability improves with model size—suggesting introspection is a learnable skill rather than an emergent property only of large models.

A common evaluation method for AI self-awareness is unreliable: The paper identifies a significant flaw in how researchers typically test whether models can detect internal perturbations. Simple yes/no questions produce misleading results in smaller models because the steering itself biases them toward saying "yes" regardless of what they actually detect, highlighting the importance of rigorous evaluation design in AI research.

Fine-tuning can dramatically improve model self-awareness: Using a technique called Introspection Fine-Tuning (IFT), researchers improved the smallest tested model's ability to detect internal changes from near-random performance (9.6% accuracy) to 60.6%—a six-fold improvement—suggesting that even limited introspective abilities can be substantially enhanced through targeted training.

Improved introspection skills transfer to related tasks: The gains from fine-tuning on one type of detection task (identifying which sentence was altered) generalized to a different but related task (comparing the strength of alterations), suggesting that introspective training builds more general self-monitoring capabilities rather than narrow task-specific skills.