Neurotechnologia
Brain-guided Language Models
2026ResearchPublished: 29 September 2026Updated: 29 September 2026Published
Key
innovation
A paradigm in which brain-activity data (fMRI) is a signal that guides a language model’s representations — not merely to measure alignment, but to actually improve reasoning robustness.
Category
Neurotechnologia
Abstraction level
Paradigm
Operation level
ModelTrainingPost-training
Use cases
Improving LLM reasoning robustnessUsing fMRI as a supervision signalNeuroAI and neural-predictivity researchSteering/fine-tuning representations (NARI/NARF)
How it works
fMRI is recorded during deductive-reasoning tasks; ridge regression builds an LLM-representation→brain mapping and identifies directions aligned with brain activity. These directions are applied as an inference-time intervention (NARI) or a supervision signal in fine-tuning (NARF), and quality is assessed via neural predictivity and reasoning-generalisation tests, among others.
Problem solved
Representational alignment between models and the brain is descriptive but does not guarantee better reasoning. This paradigm turns the brain signal into a tool that actually improves LLM robustness.
Components
fMRI data from reasoning tasksGuiding signal
Brain responses (GLMSingle betas) from selected regions.
NARI and NARF methodsParadigm implementation
Inference intervention and representation fine-tuning.
Evolution
Original paper · 2026 · Peking University (pkuxmq)
Beyond representational alignment with brain-guided language models for robust reasoning
Peking University (pkuxmq)
2026
"Beyond representational alignment" published in Nature Machine Intelligence
Inflection point