Large language models (LLMs) can now be improved by tapping into human brain signals during reasoning tasks, according to a new study from researchers at Peking University. The findings suggest that brain-guided AI may achieve more robust and human-like reasoning than language-only training alone.
The Research
Mingqing Xiao, Kai Du, and Zhouchen Lin from Peking University published their work on arXiv (June 10, 2026). They used functional MRI (fMRI) data from humans performing deductive reasoning tasks and compared these brain signals to internal representations of LLMs. Their neural-predictivity metric revealed that LLMs explain a substantial fraction of the explainable variance in reasoning-related brain regions at the aggregate level, but predictivity was lower for specific reasoning types, indicating both alignment and divergence.
The team then developed a brain-guided framework that steers LLM representations using the joint structure of model and brain data. They applied interventions at inference and fine-tuning across 10 different LLMs ranging from 1.5 billion to 72 billion parameters. The results: absolute accuracy gains of up to 13% over language-only supervision, with transfer across reasoning types. This means brain signals provide complementary information not captured by text alone.
Why It Matters
This study shows that human brain activity—specifically from reasoning regions—contains unique signals that can directly enhance machine reasoning. For cognitive science, it deepens our understanding of how reasoning is organized in the brain and how it aligns (or doesn't) with AI. For AI development, it opens a pathway toward building models that reason more robustly and align better with human cognition. For the average person, it underscores that our brain's reasoning processes are not just passive outputs but active patterns that can even teach machines.
What You Can Do
While you can't hook yourself up to an fMRI at home, you can strengthen your own reasoning skills through deliberate practice. Try solving logic puzzles, playing strategy games, or taking a free adaptive IQ test that challenges deductive thinking. The more you engage in reasoning tasks, the more you reinforce those neural circuits—and, as this research suggests, you might even help train the next generation of AI.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
