Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Ji-An, Li, Xiong, Hua-Dong, Wilson, Robert C., Mattar, Marcelo G., Benna, Marcus K.
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866908608218791936
author Ji-An, Li
Xiong, Hua-Dong
Wilson, Robert C.
Mattar, Marcelo G.
Benna, Marcus K.
author_facet Ji-An, Li
Xiong, Hua-Dong
Wilson, Robert C.
Mattar, Marcelo G.
Benna, Marcus K.
contents Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, yet at other times seem unable to recognize those strategies that govern their behavior. This suggests a limited degree of metacognition - the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognition enhances LLMs' capabilities in solving complex tasks but also raises safety concerns, as models may obfuscate their internal processes to evade neural-activation-based oversight (e.g., safety detector). Given society's increased reliance on these models, it is critical that we understand their metacognitive abilities. To address this, we introduce a neuroscience-inspired neurofeedback paradigm that uses in-context learning to quantify metacognitive abilities of LLMs to report and control their activation patterns. We demonstrate that their abilities depend on several factors: the number of in-context examples provided, the semantic interpretability of the neural activation direction (to be reported/controlled), and the variance explained by that direction. These directions span a "metacognitive space" with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a small subset of their neural activations. Our paradigm provides empirical evidence to quantify metacognition in LLMs, with significant implications for AI safety (e.g., adversarial attack and defense).
format Preprint
id arxiv_https___arxiv_org_abs_2505_13763
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations
Ji-An, Li
Xiong, Hua-Dong
Wilson, Robert C.
Mattar, Marcelo G.
Benna, Marcus K.
Artificial Intelligence
Computation and Language
Neurons and Cognition
Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, yet at other times seem unable to recognize those strategies that govern their behavior. This suggests a limited degree of metacognition - the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognition enhances LLMs' capabilities in solving complex tasks but also raises safety concerns, as models may obfuscate their internal processes to evade neural-activation-based oversight (e.g., safety detector). Given society's increased reliance on these models, it is critical that we understand their metacognitive abilities. To address this, we introduce a neuroscience-inspired neurofeedback paradigm that uses in-context learning to quantify metacognitive abilities of LLMs to report and control their activation patterns. We demonstrate that their abilities depend on several factors: the number of in-context examples provided, the semantic interpretability of the neural activation direction (to be reported/controlled), and the variance explained by that direction. These directions span a "metacognitive space" with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a small subset of their neural activations. Our paradigm provides empirical evidence to quantify metacognition in LLMs, with significant implications for AI safety (e.g., adversarial attack and defense).
title Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations
topic Artificial Intelligence
Computation and Language
Neurons and Cognition
url https://arxiv.org/abs/2505.13763