Jun-Yu Pan

Jun-Yu Pan

Incoming M.Eng. student at SJTU and Research Intern at Microsoft Research Asia

English  ·  中文

Hi, I’m Jun-Yu Pan. I will begin an M.Eng. in Electronic Information at the SJTU Paris Elite Institute of Technology, Shanghai Jiao Tong University, in September 2026, advised by Prof. Wei-Long Zheng. I received my B.Eng. from the same institute in June 2026, with an Outstanding Bachelor’s Thesis of Shanghai Jiao Tong University (Top 1%).

I am currently a Research Intern with the Brain Project (Star Project) in the Shanghai AI/ML Group at Microsoft Research Asia (MSRA), advised by Yansen Wang, and was recognized as a Top 5% Intern. My research focuses on EEG foundation models, multimodal large language models (MLLMs), and their applications to neural understanding and affective computing.

News

  • 2026.09 We developed MSDE for multi-scale future-state EEG pretraining and updated our studies on pretraining-data selection, multi-dataset fine-tuning, and generative visual grounding.
  • 2026.09 We completed a study introducing strategy-open BCI and examining how users form and revise their own interaction strategies in a continuous EEG-based task.
  • 2026.06 My bachelor’s thesis was recognized as an Outstanding Bachelor’s Thesis of Shanghai Jiao Tong University (Top 1%).
  • 2026.05 Our work on Generative Visual Grounding for EEG understanding became available on arXiv.
  • 2026.05 We completed two studies on pretraining-data scaling and multi-dataset fine-tuning for EEG foundation models.
  • 2026 We completed M4Lego, a multimodal pretraining framework for gameplay emotion recognition.
  • 2025.04 I joined Microsoft Research Asia as a Research Intern.
  • 2025 Our cross-scenario EEG emotion-recognition paper was published at ICASSP 2025.

Publications

  1. Jun-Yu Pan, Ruicheng Yin, Yansen Wang, Bao-Liang Lu, Dongsheng Li, and Wei-Long Zheng. “From Static Descriptions to State Dynamics: Multi-Scale Pretraining for EEG Transfer.” Under review, 2026.
  2. Xin Xu†, Jun-Yu Pan†, Ruicheng Yin, Zilong Wang, Dongqi Han, Dongsheng Li, Bao-Liang Lu, Wei-Long Zheng, and Yansen Wang. “Think It Your Way: Toward Strategy-Open Brain–Computer Interaction.” Under review, 2026.
  3. Jun-Yu Pan, Enze Zhang, Bao-Liang Lu, Wei-Long Zheng, Dongsheng Li, and Yansen Wang. “Visualizing the Invisible: Generative Visual Grounding Empowers Universal EEG Understanding in MLLMs.” Available on arXiv.
  4. Jun-Yu Pan, Ruicheng Yin, Bao-Liang Lu, Wei-Long Zheng, Dongsheng Li, and Yansen Wang. “Which Data, Not How Much: Architecture-Dependent Data Selection for EEG Foundation Models.” Under review, 2026.
  5. Ruicheng Yin, Jun-Yu Pan, Yansen Wang, Xiaoqing Zheng, and Dongsheng Li. “Discern, Then Combine: Probe-Guided Multi-Dataset Fine-Tuning for EEG Foundation Models.” Under review, 2026.
  6. Jun-Yu Pan†, Jing-Yi Liu†, Bao-Liang Lu, and Wei-Long Zheng. “M4Lego: A Multi-Modal Pretraining Framework with Multi-Level Masking and Lego Blocks for Emotion Recognition During Gameplay.” Under review, 2026.
  7. Jun-Yu Pan, Hao-Long Yin, and Wei-Long Zheng. “Double Domain Converter Transformer for Improving EEG-Based Emotion Recognition from Video to Game Scenarios.” ICASSP 2025.

† Equal contribution.

View all publications

Selected Research Projects

Large-Scale Pretraining and Model Scaling for EEG Foundation Models

Microsoft Research Asia · Brain Project (Star Project)

  • Investigated how increasing pretraining-data volume and model capacity affects transfer across EEG tasks, studying data and model scaling together.
  • Independently handled the full workflow, from data curation and quality control to pretraining and downstream evaluation. Expanded the corpus to 515 sources and approximately 80,000 hours of windowed EEG; conducted experiments across the 25k, 49k, and 79k corpus series and Base-to-Giant models spanning 5.8M to 1.97B encoder parameters.
  • Across 23 downstream tasks, the developed models currently rank first by a substantial margin among all EEG foundation models.

From Static Descriptions to State Dynamics: Multi-Scale Pretraining for EEG Transfer

Microsoft Research Asia · Brain Project (Star Project)

  • Studied how predicting future EEG states at different time scales can capture both transient events and sustained rhythms for transfer across tasks.
  • Developed MSDE, which jointly predicts micro- and macro-scale future teacher states using a causal encoder and a shared, elapsed-time-conditioned transition operator.
  • MSDE-Full and MSDE-10% achieved the two best average fine-tuning ranks across 20 OmniEEG-Bench tasks among eight evaluated models. MSDE-10% uses a 5.0M-parameter encoder pretrained on 2,115 hours of EEG; joint-scale supervision outperformed both single-scale variants on 18 of 20 tasks.

Think It Your Way: Toward Strategy-Open Brain–Computer Interaction

Microsoft Research Asia · Brain Project (Star Project)

  • Introduced strategy-open BCI, which specifies the task goal while allowing users to form and revise their own mental strategies rather than follow a prescribed interaction mapping.
  • Conducted an offline study with 26 participants in a continuous three-lane parkour task using EEG and self-reports. Strategies were heterogeneous and revised mainly early; reported modalities partially matched physiological signatures, while movement direction and target lane remained decodable across visual themes.

Visualizing the Invisible: Generative Visual Grounding Empowers Universal EEG Understanding in MLLMs

Microsoft Research Asia · Brain Project (Star Project)

  • Studied how MLLMs can use their pretrained visual knowledge to improve EEG understanding and generation, including tasks without visual stimuli.
  • Proposed GVG, using an EEG-to-image generator to create sample-specific proxy images, predict native visual tokens, and align EEG, images, and text.
  • On six tasks, GVG-Janus improved mean balanced accuracy from 45.89% with text-only alignment to 53.46% with image-only alignment; trimodal alignment with full adaptation reached 60.05%. GVG-X-Omni approached NeuroLM-XL’s average performance while tuning 170M parameters over a frozen 7B backbone.

Which Data, Not How Much: Architecture-Dependent Data Selection for EEG Foundation Models

Microsoft Research Asia · Brain Project (Star Project)

  • Studied interference among EEG pretraining datasets and how source composition, data volume, and model architecture affect downstream transfer.
  • Audited 202 configurations across LaBraM, CBraMod, and BIOT, covering 16 candidate datasets and 13 downstream tasks; compared Progressive, Pairwise Dataset Valuation (PDV), Size, and Coverage rankings.
  • Found 121 configurations that used fewer windows yet outperformed full-pool pretraining. The best subsets improved mean balanced accuracy by up to 5.58 percentage points; all 39 architecture-task combinations had a smaller, higher-scoring configuration.

Discern, Then Combine: Probe-Guided Multi-Dataset Fine-Tuning for EEG Foundation Models

Microsoft Research Asia · Brain Project (Star Project)

  • Studied interference during downstream multi-dataset fine-tuning and when auxiliary EEG datasets help or hurt a target task.
  • Developed POSE, which uses MMD and Fréchet-distance levels and trends during target adaptation to select and weight auxiliary datasets, then schedules their participation through warm-up, joint training, and cool-down.
  • Evaluated on 12 datasets. Mean balanced accuracy over 11 classification tasks increased from 54.5% to 56.0% on LaBraM and from 52.8% to 55.0% on CBraMod versus target-only fine-tuning, with gains on 10 and 9 of the 12 targets, respectively.

Cross-Scenario and Multimodal Emotion Recognition

BCMI, Shanghai Jiao Tong University

  • Proposed DDCT for EEG emotion recognition across video and gameplay scenarios; the work was published at ICASSP 2025.
  • Developed M4Lego, a multimodal pretraining framework integrating EEG and eye movements for emotion recognition during gameplay.

Research Experience

English CV (PDF)