AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

CerebraGloss: Instruction-Tuning a Large Vision-Language Model for Fine-Grained Clinical EEG Interpretation

Wei Gu, Luo Tianming, Qiran Zhang, Mohan Ye, Xiao Shen, Wenxin Chen, Yunhuan Li, Yichen Zhang, Jing Hong, Bao-liang Lu, Wei-Long Zheng

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Visual instruction tuning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Instruction-tunes a large vision-language model on EEG visualizations paired with fine-grained expert-style annotations to enable holistic clinical EEG interpretation.

Problem

Clinical EEG interpretation is laborious and subjective, and existing computational models only handle narrow classification tasks; LVLMs need datasets pairing EEG visualizations with fine-grained expert annotations, which are scarce.

Approach

Introduces a dataset pairing EEG visualizations with fine-grained, expert-level annotations and instruction-tunes an LVLM on it for holistic EEG interpretation rather than single-label classification.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
instruction-tune Qwen2.5-VL-3B on EEG data (CerebraGloss)Qwen2.5-VL-3B · TUSZbalanced accuracy: 55.02 -> 79.21Improved
instruction-tune on EEG data versus GPT-5CerebraGloss-3B · CerebraGloss-Bench MCQaccuracy: 70.00 -> 80.00Improved
use CerebraGloss against specialized large EEG model on sleep stagingCerebraGloss-3B · HMCbalanced accuracy: 69.97 -> 62.02Hurt
add Stage 1 alignment stopped early at 0.05 epochs (ablation)CerebraGloss · CerebraGloss-BenchMCQ/Description/QA: MCQ 78.89 -> 80.00Improved
train Stage 1 longer (0.2 epochs) (ablation)CerebraGloss · CerebraGloss-Bench MCQaccuracy: 80.00 -> 74.44Hurt
add Stage 1 alignment (ablation)CerebraGloss · TUSZ and HMCbalanced accuracy: TUSZ 79.68 vs 79.21No clear effect
add 94K LLM-generated instruction data (ablation)CerebraGloss · CerebraGloss-Bench DescriptionROUGE-1: 9.02 -> 44.19Improved
add 50K rule-based captions to Stage 2 (ablation)CerebraGloss · CerebraGloss-Bench DescriptionROUGE-1: Description 51.09 -> 44.19; MCQ 78.89 -> 80.00Mixed
scale base model from 3B to 7B (ablation)CerebraGloss · TUSZbalanced accuracy: 79.21 -> 80.21Improved
add general-domain data to mitigate forgettingCerebraGloss · MMBenchaccuracyNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy