Beyond representational alignment with brain-guided language models for robust reasoning

Large language models (LLMs) and human brain functions related to higher-order cognition have not yet been thoroughly studied in terms of correspondence. It is unclear if the neural signals from reasoning regions in the human brain can enhance LLM performance. In this research, the focus was on deductive reasoning. The study found that LLM representations are partially aligned with task-based functional magnetic resonance imaging (fMRI) activity, and these signals can directly enhance LLM reasoning.

By using a neural predictivity metric, the researchers discovered that LLMs explain a significant portion of the explainable variance in reasoning-related brain regions at an aggregate level, while predictivity within specific reasoning types is lower, indicating both alignment and divergence. Based on these findings, the authors propose a brain-guided framework that involves steering model representations along directions derived from the joint structure of model and brain representations.

This intervention is applied during both inference and training. The results demonstrate that task-evoked brain signals can directly improve LLM reasoning, leading to gains that are independent of language-only supervision across ten LLMs with parameter ranges from 1.5B to 72B. These gains were consistent across various reasoning types and resulted in up to 13% absolute accuracy improvement.

The findings bridge the gap between LLM-brain correspondences from mere correlation to guidance, paving the way for more robust and cognitively aligned artificial intelligence. Datasets utilized in this study are publicly accessible, and the implementation code is available for further research.