BrainLoRA fine-tunes up to 104B-parameter models for brain encoding
Original title: Scaling Parameter-Efficient Fine-Tuning for Brain Encoding
An ensemble of more than 200 adapted models leads the public Algonauts post-challenge leaderboard as of the paper's writing.
The paper shows parameter-efficient fine-tuning can adapt very large pretrained backbones for predicting neural responses when neural data is limited. Using the backbone's causal structure, multiple timepoints are predicted in one pass instead of sliding windows. BrainLoRA, a LoRA plus a lightweight readout, improves results across nine multimodal backbones and reports the best published score on Algonauts 2025 for the same Qwen3-Omni backbone as the prior state of the art. A multi-LoRA kernel trains many adapters on one shared backbone.
Unconfirmed: The leaderboard rank is claimed as of the time of writing and is not independently verified here.
- The largest adapted backbone has up to 104B parameters, per the authors.
- Performance improves across nine diverse multimodal backbones.
- On Qwen3-Omni it reports the best published Algonauts 2025 Challenge result.
- Adaptation increases audio contribution and cross-modal mixing inside the backbone.
- AI models
- Qwen3-Omni
From the original article
1 min read at huggingface.co
Brain 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, Guanglong Sun, Nolan Ho · huggingface.co
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.