Brain Signals Guide Language Models for Robust Reasoning, Nature Study Shows

Researchers from Peking University, Tsinghua University, and Microsoft Research Asia published a study in Nature Machine Intelligence titled "Beyond representational alignment with brain-guided language models for robust reasoning" (Xiao, Du, Lin, 2026). The paper investigates whether language models (LMs) share representational structure with human brain regions involved in deductive reasoning, and whether brain signals can functionally guide LMs to reason more robustly.

Background: Representational Alignment Between AI and Brain

The authors note that similar behavioral performance does not imply similar internal mechanisms. Representational alignment compares the structure of model hidden states with neural activity under the same stimuli. A common metric is neural predictivity (or brain score): a linear mapping is learned from model representations to fMRI responses on a training set, then used to predict held-out neural responses. The correlation between predicted and actual responses, normalized by the noise ceiling, quantifies alignment.

Partial Representational Alignment in Deductive Reasoning

The study used public task-fMRI data where participants performed syllogistic and transitive reasoning with pseudowords and controlled relational attributes to minimize semantic confounds. Ten open-source models from Qwen, Llama, Mistral, Phi, and Gemma families (1.5B–72B parameters) processed the same stimuli, and their internal activations were extracted.

Neural predictivity was computed for three brain networks: deductive reasoning regions, the multiple-demand network, and the core language network. Across all reasoning tasks, model representations explained ~76% of explainable variance in reasoning-related regions, with significantly higher predictivity in deductive reasoning and multiple-demand networks than in the core language network. This pattern held consistently across model families.

When analyzing each reasoning type separately, predictivity dropped to ~27%, indicating substantial representational differences at finer granularity. Still, trained models significantly outperformed untrained baselines and retained regional specificity. The authors conclude that models capture a partial, region-specific alignment with human reasoning representations.

From Correlation to Neural Guidance

Leveraging the partial alignment, the authors propose two neural guidance methods. First, a linear mapping from model hidden states to fMRI space is learned. For questions the model answers incorrectly but humans answer correctly, the similarity between the mapped model representation and the true fMRI response is computed, and its gradient with respect to the hidden state yields a neural guidance direction in the model's latent space.

NARI: Inference-Time Representation Intervention

Neural Activity-guided Representation Intervention (NARI) applies a constrained perturbation along the guidance direction to the attention output at a middle layer during inference, without updating parameters. On originally incorrect questions, NARI achieved 100% correction coverage, significantly outperforming random fMRI signal and random direction baselines.

Aggregating successful intervention directions across multiple questions and participants produced a transferable direction that improved performance on new reasoning questions across several models, and partially extended to chain-of-thought reasoning, suggesting the direction encodes general reasoning-relevant information.

NARF: Internalizing Guidance via Parameter Fine-Tuning

Neural Activity-guided Representation Fine-Tuning (NARF) fine-tunes model parameters so that middle-layer representations move toward the brain-guided target. The training set comprised the fMRI experiment's syllogistic and transitive problems; test sets introduced new problems, varied premise orders and counts, and different logical types (propositional reasoning) to assess robust generalization.

Fine-tuned models showed stable gains across premise counts and permutations, with clearer separation of correct vs. incorrect logic problems in middle-layer representations. Gains transferred to propositional reasoning, indicating acquisition of cross-task, cross-type representational knowledge rather than memorization.

Complementarity with Language Supervision

The authors demonstrate that neural guidance complements standard language supervision (cross-entropy on output tokens). Joint training combines the language loss (constraining final behavior) with NARF (constraining intermediate representations). Compared to language supervision alone, adding neural guidance yielded a consistent average accuracy improvement of ~2.2 percentage points, generalizing to propositional and natural-language first-order logic reasoning, with a maximum gain of 13.2 points. Benefits persisted when language supervision used synthetic training data.

Layer-wise trajectory analysis revealed that neural guidance produces more discriminative inference trajectories for correct vs. incorrect logic problems than language supervision alone, confirming it provides an additional learning signal.

Generalization to an Independent Dataset

The framework was extended to the Human Connectome Project's relational processing task. Task-fMRI activity again induced effective intervention directions, and neural guidance provided consistent gains on top of language supervision, demonstrating the approach is not limited to a single deductive reasoning dataset.

Limitations and Future Work

The authors acknowledge that current experiments focus on basic deductive reasoning, and fMRI's limited temporal resolution cannot track rapid dynamics of complex thought. Future work with higher-resolution modalities (EEG, MEG) and richer cognitive paradigms may enable finer-grained analysis and guidance.

References

[1] Fedorenko, Piantadosi & Gibson (2024). Language is primarily a tool for communication rather than thought. Nature .

[2] Schrimpf et al. (2021). The neural architecture of language: integrative modeling converges on predictive processing. PNAS .

[3] Du et al. (2025). Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence .

[4] Schrimpf et al. (2018). Brain-Score: which artificial neural network for object recognition is most brain-like? bioRxiv .

[5] Malek et al. (2025). Frontier LLMs still struggle with simple reasoning tasks. arXiv:2507.07313 .

[6] Chen et al. (2024). Premise order matters in reasoning with large language models. ICML .

Paper: https://www.nature.com/articles/s42256-026-01278-w | Preprint: https://arxiv.org/abs/2606.11893 | Code: https://github.com/pkuxmq/Brain-guided_LLM | News & Views: https://www.nature.com/articles/s42256-026-01302-z

Code example

[1] Fedorenko, E., Piantadosi, S. T. & Gibson, E. A. F. Language is primarily a tool for communication rather than thought. Nature, 2024.
[2] Schrimpf, M. et al. The neural architecture of language: integrative modeling converges on predictive processing. PNAS, 2021.
[3] Du, C. et al. Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, 2025.
[4] Schrimpf, M. et al. Brain-Score: which artificial neural network for object recognition is most brain-like? bioRxiv, 2018. https://doi.org/10.1101/407007.
[5] Malek, A. et al. Frontier LLMs still struggle with simple reasoning tasks. arXiv preprint arXiv:2507.07313.
[6] Chen, X. et al. Premise order matters in reasoning with large language models. ICML, 2024.