Brain Reasoning Models: A Framework for Extracting Brain Signatures of Mathematical Reasoning Using GNN and fMRI

The Brain Reasoning Models (BRMs) project, developed by iririthik, focuses on using Graph Neural Networks (GNN) and Meta's Tribe V2 model to extract "brain signatures" of mathematical reasoning from functional magnetic resonance imaging (fMRI) data, enabling organic arithmetic reasoning capabilities without explicit training. Unlike traditional AI mathematical reasoning which relies on large amounts of labeled data and symbolic inputs, BRMs opens up a new reasoning paradigm by simulating the neural activation patterns of the human brain.

Traditional AI mathematical reasoning usually uses explicit input combination equations, relies on external tokenizers to process expressions, and requires large amounts of labeled data for supervised learning. BRMs, however, takes a different approach: it does not directly input mathematical expressions, does not rely on external tokenizers, but instead learns through a carefully designed dictionary of mathematical building blocks, focusing on the topological overlap of brain signatures to achieve reasoning.

The core of BRMs is the Exact-Word Mathematical Model end-to-end framework, whose dictionary of mathematical building blocks includes numbers (0-9), linguistic number representations (e.g., English words), basic operators (+, -, ×, ÷, =), and functional structural words (e.g., sum, plus). The framework uses Graph Convolutional Networks (GCN) to achieve generalization, synthesis, and triggering of relevant mathematical concepts by learning the biological and topological overlap (Jaccard overlap) of brain signatures. The execution flow is divided into five steps: Vocabulary Configuration → Generate Functional Neural Activation → Build Wernicke Graph → Train Brain Signature Network → Interactive Temporal Reasoning.

The decoder of BRMs is a multi-label signature extractor. When inputting a mathematical problem, it outputs an overall probability activation distribution (e.g., "24" 0.99, "+" 0.98, etc.), which can retrieve complete concept fingerprints and activate multiple related concepts. The system training is efficient: the vocabulary is streamlined to numbers 0-100 and basic operators, training time takes less than one minute, and reasoning speed supports real-time interaction.

The application prospects of BRMs include cognitive science research (identifying key brain regions for mathematical reasoning, exploring neuroplasticity), educational applications (personalized learning paths, diagnosis of learning difficulties), and AI system enhancement (complex mathematical reasoning, logical problem solving). Current limitations: the vocabulary is limited to basic arithmetic (0-100), verified only on simple problems, and relies on simulated brain data generated by Tribe V2. Future work needs to expand to complex mathematical fields, verify with real fMRI data, and improve reasoning accuracy.

BRMs represents a brand-new paradigm for AI mathematical reasoning, which does not rely on large-scale supervised learning but instead achieves organic reasoning by simulating the neural activation patterns of the human brain. Its core insight is that mathematical reasoning is the result of the collaborative activation of multiple neural signatures in the brain. This project provides research directions for the intersection of neuroscience, cognitive science, and artificial intelligence, and is expected to promote the construction of more human-like intelligent systems.