The article is complete at 2203 words, within the target range. The file has been written to `/var/www/simpleprog-website/papers/brain-taskonomy-fmri-pretraining.html`. Here is a summary of the article:

**Title:** BrainTaskonomy: Learning How to Pretrain and What to Transfer in fMRI Foundation Models

**Structure:**

- Opens with the core problem: fMRI foundation models treat heterogeneous domains as flat mixtures and adapt tasks independently

- **The Two Allocation Problems** - Frames the pretraining and adaptation challenges

- **Measuring Domain Difficulty and Directed Facilitation** - Explains the proxy difficulty metric and compute-matched facilitation scores, including the asymmetric facilitation matrix

- **The Nested Curriculum: Domains and Timesteps** - Details the two-axis curriculum (domain ordering + high-to-low-noise timestep scheduling), the six configuration ablation, and the 6.5%/16.3%/10.5% improvements

- **The Downstream Taskonomy** - Covers first-order and higher-order transfer measurement, the asymmetric transfer structure, and the 39% vs 32% within/cross-dataset positive transfer rates

- **Budget-Aware Transfer Planning** - Explains the BIP formulation, compact source portfolios, recurrent hubs, and policy validation results

- **Pretraining Transfer to Downstream Tasks** - Reports the six-task evaluation with in-domain and out-of-domain results

- **Limitations** - Notes the exploratory validation, proxy nature of difficulty/facilitation metrics, and compute efficiency caveat

- **What This Means for fMRI Foundation Models** - Practical takeaways

**Assessment:** IMPORTANT: no - This is a solid contribution to the neuroscience/ML intersection but does not represent a major milestone or breakthrough with broad impact. It improves on existing curriculum learning and taskonomy techniques applied to fMRI, which is a niche domain.

Read the paper on arXiv