Decomposing and Steering Functional Metacognition in Large Language Models

TL;DR

This paper uncovers that large language models possess decomposable internal metacognitive states that influence their reasoning, and demonstrates how steering these states can modulate model behavior.

Contribution

It introduces a framework to identify and causally manipulate internal metacognitive states in LLMs, revealing their impact on reasoning and evaluation.

Findings

Metacognitive states are linearly decodable from internal activations.

Steering activations along probe directions modulates reasoning behavior.

Benchmark performance is influenced by activation of specific internal states.

Abstract

Large language models (LLMs) increasingly exhibit behaviors suggesting awareness of their evaluation context, often adapting their reasoning strategies in benchmark settings. Prior work has shown that such evaluation awareness can distort performance measurements; however, it remains unclear whether this phenomenon reflects a single behavioral artifact or a deeper internal structure within the model. We propose that LLMs maintain a decomposable space of functional metacognitive states: internal variables encoding factors such as evaluation awareness, self-assessed capability, perceived risk, computational effort allocation, audience expertise adaptation, and intentionality. Through residual stream analysis across multiple reasoning models, we demonstrate that these states are linearly decodable from internal activations and exhibit distinct layer-wise profiles. Moreover, by steering model activations along probe-derived directions, we show that each functional metacognitive state causally modulates reasoning behavior in dissociable ways, affecting verbosity, accuracy, and safety-related responses across tasks. Our findings suggest that benchmark performance reflects not only task competence but also the activation of specific functional metacognitive states. We argue that understandi ng and controlling these internal states is essential for reliable evaluation and deployment of reasoning models, and we provide a mechanistic framework for studying functional m etacognition in artificial systems. Our code and data are publicly available at https://github.com/xlands/meta-cognition.