Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis
arXiv Preprint Archive June 26, 2025 Jingkai Li
Applying Integrated Information Theory (IIT) 3.0 and 4.0 to sequences of large language model (LLM) representations from Theory of Mind tests reveals no statistically significant indicators of consciousness phenomena. The study compared IIT metrics—Φ^max, Φ, Conceptual Information, and Φ-structure—with Span Representations independent of consciousness estimates. Results show that contemporary Transformer-based LLM representations lack significant consciousness indicators, though spatio-permutational analyses revealed intriguing patterns.