Columbia BrainLoRA Scales PEFT to 104B Multimodal Backbones, Tops Brain-Encoding Leaderboard With 5% of Prior Params

Columbia, Tsinghua and HKUST(GZ) adapt nine multimodal backbones from 4.7B to 104B parameters for fMRI brain encoding using LoRA. A 61.2M-parameter adapter+readout matches prior 1.1B-parameter SOTA; a 200+ model ensemble reaches 0.3428 ID / 0.2483 OOD correlation, ranking first on the public leaderboard while cutting feature-extraction cost ~350x.