| name | flow-matching-in-context-brain-dynamics |
| description | Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics methodology. First generative model of whole-cortex fMRI dynamics for unseen cognitive tasks. Per-timestep conditioned diffusion transformer with compositional language priors and spatial priors for counterfactual neuroscience. |
Flow Matching with In-Context Priors for Brain Dynamics
arXiv ID: 2606.11833
Authors: Sam Gijsen, Michał Łukomski, Marc-André Schulz, Kerstin Ritter
Published: 2026-06-10
URL: https://arxiv.org/abs/2606.11833
Problem Statement
Generative models of neural time series (fMRI) have remained restricted to categorical conditioning, precluding:
- Compositional generalization
- Zero-shot generation
- Counterfactual neuroscience experiments
Existing approaches cannot generate realistic brain dynamics for unseen cognitive tasks.
Key Innovation
First generative model of whole-cortex fMRI dynamics for unseen cognitive tasks using:
- Per-timestep conditioned diffusion transformer
- Compositional language priors (in-context)
- Optional spatial priors (in-context)
Architecture
Dual Conditioning Pathway
- Language Priors: Compositional task descriptions injected in-context
- Spatial Priors: Optional region-specific activation patterns
- Synergy: Spatial priors anchor generation where language alone degrades
Diffusion Transformer
- Per-timestep conditioning
- Zero-shot generation capability
- Held-out task evaluation across hundreds of conditions
Results
Language Pathway
- Recovers region-specific recruitment across tasks
- Predicts held-out spatial activation patterns
- Compositional structure preserved
Spatial + Language
- Anchors generation in degraded regions
- Maintains compositional counterfactual specification
- Superior performance vs language alone
Applications
Counterfactual Neuroscience
- In-silico design of novel cognitive experiments
- Pre-empirical validation of hypotheses
- Data-driven experimental design
Zero-Shot Generation
- Unseen cognitive tasks from language descriptions
- Compositional task combinations
- Novel experiment prototyping
Methodology
1. Training
- Train on known cognitive tasks
- Condition per-timestep with language descriptions
- Optional spatial prior injection
2. Generation
- Language-only: zero-shot from descriptions
- Language + Spatial: anchored generation
- Evaluate across held-out task manifold
3. Validation
- Region-specific recruitment recovery
- Spatial activation pattern accuracy
- Training manifold characterization
Use Cases
- Design cognitive experiments before empirical testing
- Generate brain activity for novel task combinations
- Predict fMRI patterns for unseen tasks
- Counterfactual neuroscience simulations
Cross-Domain Connections
- Brain Dynamics: fMRI generative modeling
- Flow Matching: Diffusion transformers
- Zero-Shot Learning: Language-conditioned generation
- Counterfactual Reasoning: In-silico experiments
Activation Keywords
counterfactual neuroscience, zero-shot fMRI, flow matching brain, in-context priors, diffusion transformer brain, generative brain dynamics, language-conditioned fMRI
Code Availability
Available at: (link in arXiv comments)
