Brain-guided LLMs boost reasoning accuracy by 13% in new study
fMRI signals from human brains directly enhance LLM deductive reasoning...
A new paper from researchers at Peking University marks a significant step in bridging AI and neuroscience. The team, led by Mingqing Xiao and including Kai Du and Zhouchen Lin, demonstrates that large language models (LLMs) can not only align with human brain activity during reasoning tasks but can also be directly improved by injecting neural signals from fMRI scans.
The study focuses on deductive reasoning and finds that LLM internal representations partially correspond with task-fMRI activity in reasoning-related brain regions. Using a neural-predictivity metric, they show LLMs explain a substantial fraction of the explainable variance at the aggregate level, though predictivity varies across specific reasoning types. Building on this, they propose a brain-guided framework that steers model representations along directions derived from the joint structure of model and brain data. Applying interventions at inference and fine-tuning during training, they achieve up to 13% absolute accuracy gains across 10 LLMs ranging from 1.5B to 72B parameters, with transfer between reasoning types. This work moves beyond simple correlation to causal enhancement of AI reasoning via brain signals.
- LLMs partially align with fMRI signals from human reasoning regions, but specific reasoning types show lower alignment
- Brain-guided framework applies inference-time intervention and fine-tuning using fMRI data to steer model representations
- Achieves up to 13% absolute accuracy gain across 10 LLMs (1.5B-72B) without additional language-only supervision
Why It Matters
This opens a new pathway for building more robust, human-aligned AI by using brain activity as direct supervisory signals.