Neurotechnologia
NARI
2026ResearchPublished: 29 September 2026Updated: 29 September 2026Published
Key
innovation
A technique that intervenes on LLM representations along directions derived from brain activity (fMRI) to strengthen robust reasoning without retraining weights.
Category
Neurotechnologia
Abstraction level
Pattern
Operation level
InferencePost-training
Use cases
Strengthening robust deductive reasoningSteering LLM representations at inferenceNeuroAI research (brain as a supervision signal)Interventions without weight retraining
How it works
From model activations and fMRI responses (beta values via GLMSingle, selected regions: deductive-reasoning, language, multiple-demand networks), ridge regression builds a representation→brain mapping. Optimisation (cosine/mse/pearsonr) yields intervention directions, which are then added to hidden states (scaling 0–1.0+) at inference, steering model behaviour without changing parameters.
Problem solved
LLMs can be brittle in reasoning and diverge from human processing. NARI injects "biologically grounded" directions into representations to make reasoning more robust — without costly retraining.
Components
Representation→fMRI mapping (ridge)Deriving directions
Ridge regression linking model states to brain responses.
Intervention directionSteering
A vector in representation space added to hidden states.
Evolution
Original paper · 2026 · Peking University (pkuxmq)
Beyond representational alignment with brain-guided language models for robust reasoning
Peking University (pkuxmq)
2026
Brain-guided language models (Nature Machine Intelligence) introduce NARI
Inflection point