Detecting the Disturbance: A Nuanced View of Introspective Abilities in LLMs

Researchers investigated whether large language models can introspect by detecting perturbations to their internal states using Meta-Llama-3.1-8B-Instruct. They found that while binary detection methods from prior work were flawed due to methodological artifacts, models do show partial introspection capabilities, localizing sentence injections at 88% accuracy and discriminating injection strengths at 83% accuracy, but only for early-layer perturbations.

• →Previous binary detection paradigms for LLM introspection were confounded by global logit shifts that bias models toward affirmative responses regardless of content.
• →LLMs demonstrate partial introspection capabilities, achieving 88% accuracy in localizing which of 10 sentences received perturbations versus 10% chance.
• →Models can discriminate relative injection strengths at 83% accuracy compared to 50% chance baseline.
• →Introspection capabilities are limited to early-layer injections and collapse to chance levels for later layers.
• →The phenomenon is explained mechanistically through attention-based signal routing and residual stream recovery dynamics.