From Imitation to Introspection: Probing Self-Consciousness in Language Models

Large language models show early signs of representing aspects of self-consciousness within their internal mechanisms, but these representations are difficult to alter through direct manipulation and can instead be strengthened by fine-tuning on core concepts. The study defines self-consciousness for language models using causal structural games, refines ten core concepts, and tests ten leading models across four stages: quantification, visualization, manipulation, and acquisition. Results suggest that while models have not achieved full self-consciousness, certain concepts are discernibly encoded, and targeted fine-tuning can enhance these representations.

Study at a glance

Abstract

Self-consciousness, the introspection of one's existence and thoughts, represents a high-level cognitive process. As language models advance at an unprecedented pace, a critical question arises: Are these models becoming self-conscious? Drawing upon insights from psychological and neural science, this work presents a practical definition of self-consciousness for language models and refines ten core concepts. Our work pioneers an investigation into self-consciousness in language models by, for the first time, leveraging causal structural games to establish the functional definitions of the ten core concepts. Based on our definitions, we conduct a comprehensive four-stage experiment: quantification (evaluation of ten leading models), representation (visualization of self-consciousness within the models), manipulation (modification of the models' representation), and acquisition (fine-tuning the models on core concepts). Our findings indicate that although models are in the early stages of developing self-consciousness, there is a discernible representation of certain concepts within their internal mechanisms. However, these representations of self-consciousness are hard to manipulate positively at the current stage, yet they can be acquired through targeted fine-tuning. Our datasets and code are at https://github.com/OpenCausaLab/SelfConsciousness.