2026-08-23
Task fMRI from 10 subjects steers ten LLMs (1.5B-72B): +2.2 points over language labels on average, up to 13.2 on propositional reasoning.
LLMs can solve some logic items and still collapse when the lexicon or premise order changes. In the human brain, deductive reasoning recruits a fronto-parietal network that is largely separate from classic language cortex. Getting the answer right is not evidence that the model is running a human-like reasoning circuit.
This Nature Machine Intelligence paper, from Peking University, Tsinghua, and Microsoft Research Asia, splits the question in two. First: do open-source LLM hidden states predict task fMRI while people do syllogistic and transitive reasoning? Second: if the match is only partial, can those brain signals still supply a direction that pushes a wrong representation onto the correct side? Alignment scores are the entry ticket. The test that matters is error correction and out-of-distribution transfer.
The fMRI data are OpenNeuro ds003076. After exclusions, 10 adults aged 19-30 completed 70 items built from pseudowords and invented names, which cuts semantic-memory shortcuts. Human accuracy averaged 86.4%. Ten instruction-tuned models, Qwen2-1.5B through Llama3.3-70B, scored 71.4% to 97.1% on the same items.
Alignment is neural predictivity. Units that respond more to reasoning than to reading premises are selected, ridge regression maps them to voxels, and the correlation is divided by a human-to-human ceiling (about 0.257 in reasoning regions).
The two control knobs are NARI and NARF. Each layer learns a linear encoder W from hidden states to fMRI. On items the model misses and a subject gets right, the hidden state is moved along the gradient of Sim(Wx, y). Dropping the intercept folds in a systematic mean shift between model and brain. NARI edits attention outputs in the middle half of the stack at inference, clipped to ασ, for up to 200 steps; successful directions are averaged into a general vector. NARF either regresses toward those intervened states or adds a negative-similarity term to True/False cross-entropy, with LoRA on the same attention modules.
Across all items, the ten models explain 75.9% of explainable variance in reasoning regions versus 65.1% in language regions. Split by reasoning type, the scores fall to about 26.8% and 28.4%. Untrained weights sit well below trained models. Middle layers and attention outputs align more. The aggregate-versus-type gap is mostly type separation in representation space: a model can look well aligned just by distinguishing syllogisms from transitive items.
On six models that still err, instance-level NARI flips every miss to a hit. Random fMRI and random directions lag. Reasoning-region signals beat multiple-demand and language cortex. A pooled direction transfers to new items for some models, not for Qwen2-7B or Mistral-7B. The same edit can rewrite DeepSeek-R1-Distill-Qwen-1.5B chain-of-thought text.
Standalone NARF lifts Mistral-7B from 75.9% to 88.0% on three-premise tests with permuted order, and from 56.6% to 78.9% with six premises. Combined with language labels, the brain term adds 2.2 points on average across ten models (0.5 to 6.4), with Llama2-7B at the top of that range. On propositional items, Mistral-7B's best of five runs goes from 70.5% to 83.7%. Near 70B the ceiling is already close: Llama3.3-70B moves from 93.92% to 94.44%, and Qwen2-72B's gain is not significant. FOLIO is mixed: Qwen2-1.5B rises from 77.9% to 83.2%, Llama3-8B drops from 90.7% to 89.8%. On HCP relational reasoning, instance intervention succeeds about 80% of the time. MMLU, GSM8K, HumanEval, and MBPP stay roughly flat.
Most NeuroAI work stops at a brain score. Here, task fMRI becomes a representation-level training signal that is orthogonal to token-level labels. Language supervision mostly shapes the last layer; NARF separates correct from incorrect trajectories through intermediate layers. Without task fMRI, the pipeline does not travel. What does travel is the diagnosis: on small models and pseudoword logic, brain structure supplies a process signal that extra language labels do not. It is not a general recipe for reasoning, and it does not replace scale or data.
The 100% instance success rate means a bounded search of up to 200 steps can find a flipping direction, not that a single fixed vector works off the shelf. The same procedure on already-correct items flips 37.6% of them to wrong at α=10. fMRI is a slow hemodynamic signal and cannot track chain-of-thought steps; the authors point to MEG or EEG. Training uses 70 items and 10 subjects, with per-subject success from 52% to 86%. Standalone NARF slightly hurts Qwen2-1.5B on propositional items (47% to 45.2%), and Qwen2-7B is already at 97%. The method needs human-correct, model-wrong pairs. Gains at 70B sit on the noise edge.