Toward IIT-Inspired Consciousness in LLMs: A Reward-Based Learning Framework
arXiv Preprint Archive January 30, 2026 Hamid Reza Akbari, Mohammad Hossein Sameti, Amir M. Mansourian et al.
This paper proposes a reward-based training method for language models inspired by Integrated Information Theory (IIT), a mathematical framework for quantifying consciousness. The authors develop a reward function that measures text causality, coherence, and integration, and show that optimizing for this reward leads to more concise text generation. On out-of-domain tasks, careful tuning achieves up to a 31% reduction in output length while preserving accuracy comparable to the base model. The framework is simple, computationally efficient, requires no external data or auxiliary models, and uses a general capability-driven signal. The authors also analyze effects on confidence calibration and test-time computational scaling.