Neurotechnologia
NARF
2026ResearchPublished: 29 September 2026Updated: 29 September 2026Published
Key
innovation
Fine-tuning an LLM with brain-based supervision: fMRI-derived directions supervise the model’s attention representations (often via LoRA) to bake robust reasoning into the weights.
Category
Neurotechnologia
Abstraction level
Pattern
Operation level
Post-trainingTraining
Use cases
Durably strengthening deductive reasoningBrain-guided fine-tuning (LoRA)Generalisation to novel reasoning problemsNeuroAI research on brain-based supervision
How it works
Previously derived directions (from the representation→fMRI mapping) become a supervision target: a loss nudges the model’s attention representations to align with those directions. In the NARF+Label variant a standard language loss is added. Parameter-efficient fine-tuning (LoRA) modifies a small subset of weights, baking in the "biologically grounded" reasoning.
Problem solved
Inference-time intervention (NARI) is transient. NARF makes the benefit permanent in the weights so the model reasons more robustly even without on-the-fly steering.
Components
Representational supervisionTraining signal
A loss aligning attention representations to brain-derived directions.
NARF+Label variantMethod variant
Combines representational supervision with the label loss.
LoRA (PEFT)Fine-tuning mechanism
Parameter-efficient fine-tuning that persists the effect in weights.
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 NARF
Inflection point