Researchers demonstrate brain signals can directly guide and improve AI language model reasoning

A study by researchers Xiao, Du, and Lin demonstrates that task-evoked human brain signals can directly guide and improve the deductive reasoning of large language models. By steering internal representations during inference or fine-tuning, the researchers achieved accuracy gains of up to 13 percentage points across ten models. Meanwhile, research from IIIT-Hyderabad shows that smaller 3-billion parameter models can achieve brain alignment comparable to 14-billion parameter models.

Brain-guided AI reasoning improvement

• ▪Guiding large language models with task-evoked brain signals increased reasoning accuracy by up to 13 percentage points in the strongest reported cases.
• ▪Researchers Xiao, Du, and Lin developed a brain-guided framework that uses task-evoked human brain signals to improve the deductive reasoning performance of large language models.
• ▪The reasoning improvements from brain-guided directions transferred across different reasoning types and were orthogonal to gains from language-only supervision.

LLM deductive reasoning limitations

• ▪Tested language models showed lower neural predictivity within specific categories of reasoning than across the aggregate dataset, suggesting they may rely on shortcuts or different processing sequences.
• ▪Large language models typically trained on next-token prediction do not necessarily reproduce the distinct neural systems and computational routes the human brain uses for reasoning.

Brain-AI alignment research

• ▪Researchers used task-based functional magnetic resonance imaging tracking blood oxygenation changes to measure the representational overlap between human brain activity and large language models during reasoning tasks.
• ▪Researchers at the International Institute of Information Technology, Hyderabad, found that a 3-billion parameter language model achieves brain alignment comparable to a 14-billion parameter model.
• ▪A study by the International Institute of Information Technology, Hyderabad, showed a noticeable degradation in brain alignment when dropping from a 3-billion parameter model to a 1.5-billion parameter model.

Neural network brain-tuning methods

• ▪Brain-guided steering of language models can be applied during inference to modify processing on demand, or during training via fine-tuning to incorporate brain-derived information long-term.
• ▪The brain-guided framework steers a language model's internal state using patterns that are simultaneously meaningful in the model and predictive of neural activity.

4 sources