BrainLoRA adapts backbones up to 104B parameters for brain encoding
Original title: Scaling Parameter-Efficient Fine-Tuning for Brain Encoding
An ensemble of more than 200 LoRA adapters ranks first on the public Algonauts post-challenge leaderboard.
Researchers report that parameter-efficient fine-tuning can adapt very large pretrained backbones for brain encoding with few trainable parameters. A causal-structure trick predicts multiple timepoints in one pass, cutting feature-extraction cost and enabling models up to 104 billion parameters. BrainLoRA, using LoRA plus a lightweight readout, is the best published result on the Algonauts 2025 Challenge for the same Qwen3-Omni backbone as the prior state of the art, and a multi-LoRA kernel trains an ensemble of more than 200 adapters.
Unconfirmed: The 104-billion-parameter scale is described as the largest brain-encoding adaptation to the authors' knowledge.
- The method improves encoding across nine multimodal backbones.
- The largest adapted backbone has 104 billion parameters.
- Modality attribution recovers canonical sensory topography after adaptation.
- Neural adaptation increases audio and cross-modal contributions inside the backbone.
- A multi-LoRA kernel jointly trains many adapters on one shared backbone.
From the original article
1 min read at arxiv.orgBrain encoding models aim to predict neural responses to stimuli. The most widely used approach fits readouts from frozen pretrained backbones. Using neural data to train large models from scratch or fully fine-tune them after task-pretraining has proven difficult because neural data is costly and limited.
By Zhuofan Josh Ying · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.