Brain-guided steering lifts reasoning by 13 points

A brain-guided representation-steering framework produced up to a 13-percentage-point absolute accuracy gain on deductive reasoning across 10 language models ranging from 1.5 billion to 72 billion parameters. The peer-reviewed Nature Machine Intelligence paper, published August 3, applies directions derived from joint model and task-fMRI representations during inference and fine-tuning, while the open preprint reports that gains transfer across reasoning types and remain distinct from language-only supervision. For builders evaluating the frontier contest described in OpenAI’s agent-model race, this is an early research result—not a production recipe—but evidence that external cognitive signals may become another lever alongside data, reinforcement learning, and test-time compute.