| name | flow-matching-in-context-brain-dynamics |
| description | Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics methodology — per-timestep conditioned diffusion transformer for generating realistic fMRI during unseen cognitive tasks using compositional language and spatial priors. |
| version | 1.0.0 |
| category | neuroscience |
| tags | ["brain-dynamics","fMRI","flow-matching","diffusion","transformer","zero-shot","counterfactual-neuroscience","generative-model"] |
| arxiv | 2606.11833 |
| authors | ["Sam Gijsen","Michał Łukomski","Marc-André Schulz","Kerstin Ritter"] |
| published | 2026-06-10T00:00:00.000Z |
| activation_keywords | ["flow matching","brain dynamics","fMRI generation","zero-shot","in-context prior","diffusion transformer","counterfactual neuroscience","cognitive task"] |
Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics
Overview
First generative model of whole-cortex fMRI dynamics for unseen cognitive tasks, enabling counterfactual neuroscience through compositional language and spatial priors.
arXiv: 2606.11833
Authors: Sam Gijsen, Michał Łukomski, Marc-André Schulz, Kerstin Ritter
Published: June 10, 2026
Categories: cs.LG, q-bio.NC
Core Innovation
Per-timestep conditioned diffusion transformer that:
- Generates realistic fMRI brain dynamics during unseen cognitive tasks
- Injects compositional language priors for task specification
- Optionally incorporates spatial priors for region-specific anchoring
- Enables zero-shot generation for novel experimental designs
Key Methodology
1. Architecture
- Diffusion transformer backbone
- Per-timestep conditioning (not categorical)
- Dual pathway: language + optional spatial
2. In-Context Priors
- Language pathway: Task descriptions as compositional priors
- Spatial pathway: Optional ROI-level activation patterns
- Joint conditioning: Contextual integration for unseen tasks
3. Training Strategy
- Flow matching objective
- Context injection at each timestep
- Compositional generalization via language embedding
Applications
Counterfactual Neuroscience
- In-silico design of novel cognitive experiments
- Evaluate hypothetical experiments before empirical validation
- Predict brain responses to unseen task combinations
Zero-Shot Brain Dynamics Generation
- Generate fMRI for tasks outside training distribution
- Compositional task specification via natural language
- Region-specific recruitment prediction
Data-Driven Experimental Design
- Simulate cognitive experiments computationally
- Optimize experimental parameters before running
- Reduce empirical validation costs
Technical Details
Input Modalities
- Task language: Natural language task descriptions
