Nature Machine Intelligence｜Human brain signals can directly guide language models for robust reasoning

Whether language models and the human brain reason in similar ways has long remained at the level of "representational alignment": under the same stimuli or tasks, model hidden states and brain activity are extracted separately, and then the two types of representational structures are compared for consistency. The commonly used metric is neural predictivity (also called brain score), which uses model representations to predict brain activity and measures the correspondence by the degree of correlation between the predicted results and the real neural responses.

The background view cited in the report points out that the core brain network supporting language processing and the systems supporting complex thinking such as reasoning can be separated from each other; language is closer to an information transmission system serving communication, rather than a necessary condition for complex thinking. Existing NeuroAI research has found that language models' internal representations have a certain correspondence with human brain language network activity, and multimodal large models can also form structures partially similar to human object concept representations, but for higher-order tasks such as logical reasoning that go beyond language processing itself, systematic evidence is still lacking.

The open question arising from this is: even if models and the human brain have correlational alignment, does this correspondence have functional significance, that is, can human brain signals in turn be used to guide models toward robust reasoning. The study claims to advance this relationship from correlational representational analysis to functional neural guidance.

The study used public task-state fMRI data, in which participants completed two types of tasks—syllogistic reasoning and transitive relation reasoning—and judged whether the conclusion could be inferred from the premises; to reduce semantic interference, the problems were mainly composed of pseudowords, fictional names, and controlled attribute relations. On the model side, 10 open-source models from the Qwen, Llama, Mistral, Phi, and Gemma series were selected, with parameter scales covering 1.5B to 72B, and internal activation representations were extracted.

The results show that when considering the overall reasoning task, model representations on average explain about 76% of the explainable variance in reasoning-related brain regions, and there is regional specificity: neural predictivity in the deductive reasoning brain regions and the multiple-demand network is significantly higher than in the core language network, and this trend remains consistent across different model families.

But the report also gives counter-evidence: when the two types of reasoning problems are analyzed separately, neural predictivity drops markedly to about 27%, indicating that differences between model and human brain representations still exist at a more fine-grained level. Even so, the models' neural predictivity is still significantly higher than that of untrained models and maintains regional specificity. The report's conclusion is "partial alignment"—the models capture structures partially shared with human brain reasoning activity and with brain-region specificity, but there are still obvious differences within specific reasoning types.

The neural guidance method proposed in the study first learns a linear mapping from model hidden states to fMRI space, then, for problems that the model answers incorrectly while humans answer correctly, calculates a similarity metric between the mapped model representations and the real brain activity, and computes gradients with respect to the model hidden states to obtain a neural guidance direction in the model's latent space. This direction is determined by the joint structure of model representations and brain representations.

The first usage is neural activity-guided representation intervention (NARI): during model inference, it directly modifies internal representations without updating parameters. The report claims that for the original error problems included in the analysis, NARI ultimately achieves 100% error-correction coverage, significantly outperforming the random brain signal baseline that retains only the model representation structure and the random direction baseline. After further aggregating successful intervention directions generated from multiple problems and participants, a general direction transferable to new problems can be obtained, improving test performance on new reasoning problems across multiple models, and can to some extent extend to model reasoning processes with chain-of-thought.

The second usage is neural activity-guided representation fine-tuning (NARF): through fine-tuning, the model's intermediate representations are brought closer to the brain-signal-guided target, internalizing representational-level knowledge into the parameters. The report cites existing research pointing out that language models still have robust generalization problems even for relatively simple reasoning problems; for example, the order in which premises are presented can significantly affect model performance. Therefore, training uses the syllogistic and transitive relation problems from the fMRI experiment, while testing introduces new problems, different premise orders and quantities, and reasoning problems of different logical types. The results show that the fine-tuned model achieves stable improvements under different numbers and arrangements of premises, the separation of correct and incorrect logical problems in the intermediate layers becomes clearer, and the performance gain can also transfer to propositional reasoning.

The study further claims that the gains brought by neural guidance are complementary to conventional language-level supervision. Language supervision directly constrains the model to give correct answers through the cross-entropy loss on the final output tokens, while NARF simultaneously pushes intermediate representations to change in a direction closer to human task-state neural activity; the two objectives act on final behavior and internal representations respectively, so they can participate in training together.

The results given in the report are: after adding brain signal guidance, model accuracy improves consistently compared with using only language supervision, by an average of about 2.2 percentage points, and the improvement can generalize to propositional reasoning with different logical structures as well as natural language first-order logic reasoning, with a maximum improvement of up to 13.2 percentage points. The study also tested a setting using synthetic training data for language supervision, and the results show that neural guidance still provides additional gains; hierarchical trajectory analysis of internal representations shows that, compared with using only language supervision, neural guidance can form a more discriminative inference trajectory for the intermediate representations of correct and incorrect logical problems.

The framework was also extended to the relational processing task of the Human Connectome Project; the report claims that task-state brain activity can also induce effective representation intervention directions, and neural guidance can also bring consistent gains on top of language supervision, indicating that the framework is not limited to a single deductive reasoning dataset.

The report explicitly lists the limitations of the current experiments: it mainly focuses on basic deductive reasoning, and the temporal resolution of fMRI makes it difficult to track rapid changes in complex thinking processes. The authors propose that future neural data from modalities with higher temporal resolution, such as EEG and MEG, as well as neural data covering more complex cognitive processes, may support more fine-grained analysis and guidance.

It is necessary to distinguish the levels of evidence: the above numbers—76%, 27%, 100% error-correction coverage, 2.2 and 13.2 percentage points—all come from that WeChat public account's retelling of the Nature Machine Intelligence paper and are secondhand reporting; this article has not obtained the original first-party paper for verification, so these numbers and conclusions should be regarded as source claims rather than independently confirmed facts.

Also belonging to source claims are explanatory judgments such as "the brain-signal-guided direction contains more general reasoning-related information" and "what the model learns is representational knowledge across problems and reasoning types." These judgments are supported by the experimental phenomena in the report, but their generality and mechanisms still await more independent verification.