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

The study presents a brain-guided framework that enhances large language models (LLMs) by aligning their internal representations with neural signals from reasoning-related brain regions. By applying intervention during inference and fine-tuning during training, the researchers demonstrated improvements in deductive reasoning across 10 LLMs ranging from 1.5B to 72B parameters, achieving up to a 13% absolute accuracy gain. This approach shifts the focus from mere correlation between LLMs and brain activity to a more robust guidance mechanism, potentially leading to AI systems that better mimic human cognitive processes.