Toward IIT-Inspired Consciousness in LLMs: A Reward-Based Learning Framework

TL;DR

The paper addresses whether IIT concepts can induce consciousness-like processing in LLMs through reward-based learning. It introduces an Intrinsic Information (II) reward, alongside a $\Phi$-based reward, to train autoregressive policies via RL, with per-token consciousness signals derived from a binarized TPM built from token embeddings. Empirical results show significant length reductions (up to 31% on out-of-domain tasks) with accuracy that remains competitive, and favorable changes in entropy and calibration under certain settings. The work offers a simple, data-free, and scalable approach to compressing outputs while retaining performance, highlighting practical benefits for real-time deployments and informing future explorations of consciousness-inspired signals in language models.

Abstract

The pursuit of Artificial General Intelligence (AGI) is a central goal in language model development, in which consciousness-like processing could serve as a key facilitator. While current language models are not conscious, they exhibit behaviors analogous to certain aspects of consciousness. This paper investigates the implementation of a leading theory of consciousness, Integrated Information Theory (IIT), within language models via a reward-based learning paradigm. IIT provides a formal, axiom-based mathematical framework for quantifying consciousness. Drawing inspiration from its core principles, we formulate a novel reward function that quantifies a text's causality, coherence and integration, characteristics associated with conscious processing. Empirically, it is found that optimizing for this IIT-inspired 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 levels comparable to the base model. In addition to primary task performance, the broader effects of this training methodology on the model's confidence calibration and test-time computational scaling is analyzed. The proposed framework offers significant practical advantages: it is conceptually simple, computationally efficient, requires no external data or auxiliary models, and leverages a general, capability-driven signal rather than task-specific heuristics. Code available at https://github.com/MH-Sameti/LLM_PostTraining.git

Figures (8)

• Figure 1: An example showing that a language model inspired by consciousness theory needs fewer reasoning steps to reach a correct final output. Baseline refers to the model post-trained using accuracy as the reward.
• Figure 2: IIT-based reward computation. (a) TPM is constructed from the input sequence x and the generated output y. The representation of y conditioned on x is first refined through an attention-based mechanism. Dimensionality reduction is then applied using Principal Component Analysis (PCA) to project the representation onto a set of units consistent with the IIT framework. The resulting representations are subsequently binarized using mean-based thresholding. (b) Based on the TPM derived from the token sequence, the Intrinsic Information Cause ($ii_{cause}$) and Intrinsic Information Effect ($ii_{effect}$) are computed for each token. The final reward assigned to the LLM output is obtained by summing the $ii_{cause}$ and $ii_{effect}$ across all tokens.
• Figure 3: comparison of the proposed method with Baseline 1, in which the model is trained using only an accuracy-based reward. The comparison is conducted on the Open-Thought dataset across different training steps. Evaluation metrics include accuracy,entropy, an the average number of generated tokens.
• Figure 4: Phi-based reward Training
• Figure 5: II Reward Training Using Trajectory Mode