Beyond representational alignment with brain-guided language models for robust reasoning

Abstract: The correspondence between LLMs and the neural mechanisms underlying human higher-order cognition remains insufficiently characterized. Given that language and reasoning in the human brain appear dissociable, an open question is whether LLMs align with neural signals from reasoning-related regions and whether such signals can improve them. Here, focusing on deductive reasoning, we show that LLM internal representations are not only partially aligned with task-fMRI activity but can also be directly enhanced by these signals. Using a neural-predictivity metric, we find that LLMs explain a substantial fraction of the explainable variance in reasoning-related regions at the aggregate level, whereas predictivity within specific reasoning types is lower, indicating both alignment and divergence. Building on this, we propose a brain-guided framework: we steer model representations along directions induced by the joint structure of model and brain representations, applying intervention at inference and fine-tuning during training. We demonstrate that task-evoked brain signals can directly enhance LLM reasoning, yielding gains orthogonal to language-only supervision across 10 LLMs (1.5B-72B), with transfer across reasoning types and up to 13\% absolute accuracy gain. Our results advance LLM-brain correspondences from correlation to guidance, establishing a brain-signal-driven pathway toward more robust and cognitively aligned AI.

• The paper demonstrates that using fMRI signals to steer LLM latent representations via NARI and NARF achieves 100% correction on deductive errors.
• It employs ridge regression to map LLM activations to human brain regions, revealing that middle layers in reasoning ROIs are most predictive.
• Neural-guided interventions boost overall accuracy by up to 13% without compromising general-purpose language performance.

Brain-Guided LLMs for Robust Reasoning

Introduction

This paper presents a systematic investigation into the representational correspondence between LLMs and the neural mechanisms underlying human deductive reasoning, moving beyond mere behavioral alignment toward direct representational enhancement using human neural data. The authors quantify the degree of shared representational geometry between LLMs and human brains in the context of deductive reasoning and demonstrate that neural signals can be used to functionally steer and fine-tune LLM reasoning, complementing conventional language-based supervision. The experimental evidence spans both inference-time intervention in LLMs’ latent spaces and parameter fine-tuning, employing fMRI signals from human participants performing formal reasoning tasks.

Figure 1: Overview of a framework linking human and model behavior and internal representations in deductive tasks; the paper moves from behavioral alignment toward leveraging internal neural representations to guide model optimization.

Model-Brain Correspondence in Deductive Reasoning

LLMs exhibit above-chance behavioral performance on deductive tasks, but the key question is whether their internal representations are isomorphic with human neural activity in regions mediating formal reasoning. Comprehensive fMRI data (OpenNeuro ds003076) are used to directly benchmark ten diverse instruction-tuned LLMs (1.5B–72B parameters) against human neural responses elicited during syllogistic and transitive reasoning tasks involving pseudo-lexical content to minimize semantic confounds.

Using a rigorous neural predictivity framework, LLM latent activations (from localized, functionally responsive units per layer) are mapped onto multivoxel fMRI data via cross-validated ridge regression. The resulting neural predictivity scores, normalized to the explainable variance ceiling, indicate that LLMs collectively capture approximately 76% of explainable variance in deductive reasoning ROIs, vastly exceeding untrained model baselines and showing significant regional specificity: predictivity is highest in reasoning and MD (multiple demand) regions and is much lower in classical language areas.

Notably, within-reasoning-type predictivity drops to ∼27% but remains robust and well above random controls, while PCA visualizations confirm partial separation of syllogistic and transitive representations in both LLMs and fMRI, reflecting shared intermediate geometry. Layer-level analysis indicates that middle layers and attention outputs are most predictive.

Figure 2: LLM–brain neural predictivity pipeline and scoring details; higher scores for reasoning ROIs; robust relative to untrained models and language network regions.

Steering LLM Reasoning via Human Neural Activation

The partial representational overlap observed motivates a move beyond correlational alignment. The authors develop two procedures:

1. NARI (Neural Activation guided Representation Intervention): At inference time, for cases where an LLM produces an incorrect answer on a reasoning problem, fMRI signals from correct human responders are used to induce a gradient-based direction in the LLM’s intermediate representation space (focused on attention module outputs in the middle transformer layers). Projected gradient modification along this direction is performed within a bounded range, and effectiveness is measured by whether this shifts output to the correct answer. Baselines include purely random fMRI signals and random direction vectors in latent space.
2. NARF (Neural Activation guided Representation Fine-tuning):

Model parameters are fine-tuned to move latent representations (for incorrect responses) toward their NARI-induced neural targets, minimizing representation-level MSE losses, with regularization to maintain stability on already correct instances.

Figure 3: Pipeline for inducing neural-guided directions in latent space, enabling inference-time correction (NARI) and parameter fine-tuning (NARF), and integration with output-based supervision.

Figure 4: NARI representation intervention corrects model outputs on initially incorrect cases with 100% success across several model families; guided directions outperform both random signals and random directions, and are most effective when computed from reasoning ROIs.

NARI achieves 100% correction of erroneous LLM outputs on deductive test problems, with directional effectiveness persisting under narrow perturbation constraints—demonstrating a functional, not merely correlational, neural linkage. Only signals from the reasoning network—not MD or language regions—yield this effect. The induced guidance directions generalize (at inference) to new, out-of-set reasoning problems as a fixed, model-averaged vector, again yielding nontrivial accuracy gains, while random directions fail.

Intervention routes incorrect representations to correct regions in latent space, as confirmed by PCA projections pre/post-intervention. The method extends to chain-of-thought-capable (CoT) models, showing the intervention modulates entire language-based reasoning trajectories toward correct logical conclusions.

Parameter Fine-Tuning with Neural Guidance

NARF parameter updates internalize the utility of neural-guided directions, yielding reasoning gains that reliably generalize to new problems, different premise orders, and reasoning types. As shown in fine-tuning evaluations, testing accuracy improves consistently on extrapolated cases—indicating that the neural intervention does not merely memorizing training instances but imparts more robust, generalizable reasoning mechanics.

Figure 5: Generalization performance of LLMs fine-tuned using NARF on permuted-order, more-premise, and novel-type reasoning problems; fine-tuned representations better separate correct from incorrect logical outcomes.

When combined with canonical language-output supervision (i.e., adding neural activation similarity loss to standard cross-entropy decoding objectives), neural guidance yields nontrivial complementary gains—up to 13% absolute accuracy improvement in some models/reasoning types. Ablation studies show these gains cannot be attributed to random signals or label/propositional structure alone: the effect is unique to true brain structure and the joint geometry of model/brain representations.

Layerwise trajectory analysis corroborates a functional induction: hybrid models trained with neural guidance show more linearly separable and interpretable transitions between correct/incorrect logic in their deep layers—an internal “reasoning path” effect that is not observed in label-only fine-tuning.

Figure 6: Neural-guided (NARF+Label) fine-tuning outperforms label-only, with generalization to previously unseen deductive types and natural-language FOLIO problems. Layerwise reasoning trajectories become more robustly distinguished between logical outcomes.

Robustness, Extension, and Mechanistic Analysis

The representational structures of both human neural activations and LLM representations are not isomorphic across subjects or models, but the intervention’s effectiveness scales with the number and quality of human data contributors; high-performing subjects yield the highest enhancement, but NARF is robust to subject variability.

Detailed ablations, including intercept inclusion/exclusion and alternatives to cosine similarity, confirm the importance of incorporating both the geometry and systematic shift in aligning neural and model spaces for best effect.

Furthermore, the approach generalizes to other neural datasets beyond deductive reasoning, as shown in transfer experiments with the HCP Relational Processing Task, leveraging models to reason about analogously-structured image attributes. The mechanism is robust to the modality and task structure.

The paper further demonstrates that neural guidance does not induce catastrophic forgetting on general-purpose language or coding benchmarks post-fine-tuning, and neural supervision yields orthogonal improvements even in the presence of additional synthetic training data.

Implications and Future Directions

This study demonstrates that cognitive-neuroscience-based representational guidance can provide a functional augmentation to LLMs, orthogonal to output-based language supervision. The neural signals are not merely labeling proxies but encode geometric and type-specific structure that, when mapped into LLM representations, repair nontrivial model errors and induce more robust deductive computation.

This work moves beyond previous neuro-inspired AI paradigms by showing that direct functional guidance using measured human brain data enhances high-level cognitive capabilities in models at the level of internal process, not just output behavior. The results call into question the sufficiency of pure language-based supervision for robust, general reasoning, indicating that fMRI or related neural modalities can serve as a highly informative supervision signal for future cognitive alignment and reasoning-robust AI development.

From a theoretical standpoint, the findings support the existence of a partially shared representational substrate across artificial and biological systems for deductive reasoning, while also exposing their divergence—particularly at the granularity of reasoning subtype and internal layer. Practically, the methods developed here—NARI and NARF—provide ready tools to enhance model robustness, and their success with mid-sized models suggests tangible gains for efficiency and generalization that complement scaling approaches.

Conclusion

This paper establishes, with strong empirical evidence, that LLMs and human brains exhibit measurable, region-specific representational alignment during deductive reasoning and that human neural activations can be leveraged not only to diagnose but to causally enhance LLM performance on reasoning tasks. Neural-signal-guided training yields statistically significant gains that are robust, generalizable, and compatible with conventional supervision—highlighting a new process supervision axis for cognitive AI. As more temporally and structurally detailed neural data become available (e.g., MEG, EEG), such neuroscientific supervision may yield further step improvements in the alignment and robustness of AI cognitive capabilities.

(2606.11893)