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

Researchers demonstrate that large language models can be enhanced by integrating brain signals from human reasoning regions, achieving up to 13% accuracy gains on deductive reasoning tasks. By aligning LLM representations with fMRI data from reasoning-related brain regions, the study establishes a framework that guides model behavior beyond traditional language supervision alone.

This research bridges neuroscience and AI by establishing that human brain signals can directly improve language model reasoning capabilities. The study moves beyond observing correlations between LLMs and neural activity to actively using brain data as a training signal. Researchers found that while LLMs partially align with fMRI activity in reasoning regions, gaps exist in task-specific reasoning, suggesting current models diverge from human cognitive mechanisms in meaningful ways.

The work addresses a fundamental question in AI development: whether machines can benefit from reverse-engineering human intelligence. The brain-guided framework steers model representations using directions derived from joint model-brain structure, applied both during inference and training. Testing across 10 models ranging from 1.5B to 72B parameters demonstrates scalability and consistency of improvements, with gains transferring across different reasoning types.

This approach has implications for AI robustness and alignment. Rather than relying solely on traditional supervised learning or reinforcement learning from human feedback, incorporating neural signals provides an additional optimization dimension. The orthogonal improvements to language-only supervision suggest brain guidance captures aspects of reasoning that standard training methods miss. The methodology could inform future development of more cognitively aligned AI systems that better mirror human problem-solving.

Future work likely explores whether similar brain-guided approaches enhance other cognitive capabilities beyond reasoning, and whether scaling this method improves performance on complex reasoning benchmarks. The research also raises questions about which brain regions provide the most valuable guidance for different AI tasks.

• →Brain-guided fine-tuning improves LLM reasoning by up to 13% accuracy across model sizes.
• →LLM representations partially align with human brain activity in reasoning regions but show task-specific divergence.
• →Brain signals provide optimization orthogonal to traditional language supervision, suggesting novel training pathways.
• →The framework demonstrates consistent improvements across 10 models from 1.5B to 72B parameters.
• →This approach establishes brain data as actionable guidance rather than passive correlation markers for AI development.