Me, Myself, and $Ï$ : Evaluating and Explaining LLM Introspection
This work addresses the challenge of distinguishing true meta-cognition from text-based simulation in LLMs, which is incremental in advancing evaluation methods for AI introspection.
The paper tackled the problem of evaluating genuine introspection in large language models by proposing a taxonomy and Introspect-Bench, a rigorous evaluation suite, finding that frontier models outperform peers in predicting their own behavior with privileged access to their policies.
A hallmark of human intelligence is Introspection-the ability to assess and reason about one's own cognitive processes. Introspection has emerged as a promising but contested capability in large language models (LLMs). However, current evaluations often fail to distinguish genuine meta-cognition from the mere application of general world knowledge or text-based self-simulation. In this work, we propose a principled taxonomy that formalizes introspection as the latent computation of specific operators over a model's policy and parameters. To isolate the components of generalized introspection, we present Introspect-Bench, a multifaceted evaluation suite designed for rigorous capability testing. Our results show that frontier models exhibit privileged access to their own policies, outperforming peer models in predicting their own behavior. Furthermore, we provide causal, mechanistic evidence explaining both how LLMs learn to introspect without explicit training, and how the mechanism of introspection emerges via attention diffusion.