Research map / Large language models / Language, safety and interpretability
BrainCoT: A Multi-Task Zero-Shot Brain Signal Foundation Model with Neurometric-Anchored Chain-of-Thought Reasoning
Yiyu Gui, Mingzhi Chen, Guibo Luo, Yuchao Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BrainCoT, a multi-task zero-shot brain signal foundation model that uses neurometric-anchored chain-of-thought reasoning to generalize across tasks and subjects without target-specific supervised adaptation.
Problem
Brain signal models struggle to generalize across multi-center heterogeneity, inter-subject variability, and diverse tasks, and existing brain foundation models still require target-data-specific supervised adaptation.
Approach
Combines a brain signal foundation model with chain-of-thought reasoning anchored to neurometrics so that it can perform multiple tasks zero-shot without target-specific supervised fine-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| generate neurometric-anchored reasoning instead of answering directly (ablation) | BrainCoT stage-II · 12 datasets zero-shot | ACC / AUC: 62.91 -> 67.41 ACC; 69.56 -> 74.33 AUC | Improved |
| add stage-II reasoning pretraining (direct mode) (ablation) | BrainCoT · 12 datasets zero-shot | ACC / AUC: 57.59 -> 62.91 ACC; 63.18 -> 69.56 AUC | Improved |
| corrupt reasoning trajectory by randomly replacing neurometric names and intensities (ablation) | BrainCoT reasoning · 12 datasets zero-shot | ACC / AUC: 67.41 -> 60.71 ACC; 74.33 -> 67.96 AUC | Hurt |
| add Decision-Consistent Reasoning (DCR) training (ablation) | BrainCoT · 12 datasets zero-shot | accuracy | Improved |
| add MoE backbone (ablation) | BrainCoT · 12 datasets zero-shot | accuracy | Improved |
| use MI-weighted evidence sampling instead of random features (ablation) | BrainCoT · 12 datasets zero-shot | accuracy | Improved |
| scale model from Small (83.8M) to Base (329.2M) (ablation) | BrainCoT · zero-shot | ACC / AUC | Improved |
| zero-shot BrainCoT versus 1% linear probing baselines | BrainCoT · 12 datasets | ACC / AUC: 63.56 -> 67.41 ACC; 68.36 -> 74.33 AUC (strongest baselines) | Improved |
| compare BrainCoT zero-shot with 10% linear probing baselines | BrainCoT · 12 datasets | ACC: 67.41 vs 67.69 (BrainBERT) | Mixed |
| use BrainCoT instead of baselines under full fine-tuning | BrainCoT · 12 datasets | AUC: 85.01 vs 85.95 (ELM-MIL) | No clear effect |
Closest papers by meaning
- BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.87)
- Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis (ICLR 2025 · similarity 0.86)
- NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals (ICLR 2025 · similarity 0.86)
- How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding (AAAI 2026 · similarity 0.85)
- Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement (NeurIPS 2025 · similarity 0.85)
- Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
