What it is
LoRA-adapts huge multimodal backbones for brain encoding, making fine-tuning practical with single-pass causal extraction and multi-LoRA training. Instead of frozen backbones plus giant readouts, it PEFT-adapts pretrained multimodal models with small LoRAs; causal single-pass extraction replaces sliding windows, enabling adaptation up to 104B params and large adapter ensembles.
Key results
- 0.327 ID and 0.228 OOD, beating MIRAGE's 0.310/0.217 single model
- 261-model ensemble scores 0.343 ID and 0.248 OOD, top public leaderboard
- Single-pass extraction cuts feature cost about 350x vs prior sliding-window setup
Shows small neural-data adapters can beat frozen-backbone systems with much larger readouts for whole-brain encoding.