cv

PhD student

Northeastern University, Boston, US

Computer Engineering

• GPA: 3.92/4
• Advisor: Prof. Yanzhi Wang

Master of Science

King Abdullah University of Science and Technology

Computer Science

• GPA: 3.71/4
• Advisor: Marco Canini

Bachelor of Engineering

Zhejiang University

Electronic Engineering

• GPA: 3.75/4

Research Intern

Futurewei Technologies, Inc.

Research Intern, Generative AI & World Modeling team.

• Built LEAP end to end on Qwen3-Omni-30B-A3B for hour-scale audio-visual QA: the localization and answer stages, their two LoRAs, and the evaluation harness behind all reported numbers.
• Method: per-block localization picks evidence windows; one bounded answer pass re-encodes only those windows.
• Results: +4.5–16.8 pp over the Qwen3-Omni baseline on TraceAV, LVOmni, VideoOdyssey (1–4 h) and MMOU; transfers to MiniCPM-o 4.5 (+3.1–13.0 pp over its published results); answer input and peak context stay independent of recording duration; supports causal/streaming queries without streaming-specific training.

Infrastructure Engineering Intern

Beijing VirtAI Tech

• Integrated the PaddlePaddle distributed training library and Deep Graph Library (DGL) into the in-house performance measurement system, extending it to two new framework backends.
• Profiled PaddlePaddle distributed training and reported its scaling bottlenecks to the platform team.

Physical Reasoning Benchmark for Generative World Models

NEU, Sept 2025 – Mar 2026

• Proposed and built PhyGround, a criteria-grounded benchmark for physical reasoning in generative video world models built on 250 curated prompts and a taxonomy of 13 physical laws (mechanics, fluids, optics); evaluated 8 video models against it.
• Led a large-scale human annotation study: 459 annotators, 5,796 complete annotations, 37.4K+ fine-grained labels, with split-half ranking correlation Spearman's ρ > 0.90.
• Trained and released PhyJudge-9B, an open physics-specialized VLM judge reaching 3.3% aggregate relative bias vs. 16.6% for Gemini-3.1-Pro.

Video Reasoning Benchmark

NEU, Sept 2025 – Jan 2026

• Proposed and co-led (team of 22) VBVR-EvalKit, a reasoning benchmark for video generation models, evaluating 9 models (4 proprietary, 5 open-weight).
• Designed the data-generation and benchmark framework, set the roadmap, and contributed core code and evaluation metrics; shipped an open-source ecosystem with 100+ synthetic data generators.
• Built a dataset of 200 curated reasoning tasks and 1M+ video clips, 1000× larger than prior datasets; released eval suite = 100 tasks × 5 samples (500 videos).

Vision Language Action (VLA) Model

NEU & Funded by EmbodyX, Jan 2025 – Sept 2025

• Proposed VOTE, a fine-tuning framework for parallel VLA action prediction, using an ensemble voting strategy over sampled actions to cut compute while improving success and generalization; adopted by Cisco and EmbodyX.
• +20% average success over OpenVLA across four LIBERO suites, +7% over the state-of-the-art VLA model on SimplerEnv WidowX, and 39× faster action generation on NVIDIA Jetson Orin.
• Pretrained Llama3.2-1B into a VLM and fine-tuned it with VOTE, beating 7B OpenVLA by 16% average success across the four LIBERO suites.

PhyGround: Benchmarking Physical Reasoning in Generative World Models

Preprint (arXiv:2605.10806)

Juyi Lin, Arash Akbari, Yumei He, Lin Zhao, Haichao Zhang, Arman Akbari, Xingchen Xu, Zoe Y. Lu, Enfu Nan, Hokin Deng, Edmund Yeh, Sarah Ostadabbas, Yun Fu, Jennifer Dy, Pu Zhao, Yanzhi Wang

VOTE: Vision-Language-Action Model Optimization for Real-Time Robotic Manipulation

ICANN 2026

Juyi Lin, Amir Taherin, Arash Akbari, Arman Akbari, Lei Lu, Guangyu Chen, Taskin Padir, Xiaomeng Yang, Jayanth Srinivasa, Xue Lin, David Kaeli, Yanzhi Wang, Pu Zhao, Xuan Shen

A Very Big Video Reasoning Suite

ICML 2026

Maijunxian Wang*, Ruisi Wang*, Juyi Lin*, Ran Ji*, Thaddäus Wiedemer, Qingying Gao, and others (* equal contribution)

Cross-Platform Scaling of Vision-Language-Action Models from Edge to Cloud GPUs

GLSVLSI 2026

Amir Taherin, Juyi Lin, Arash Akbari, Arman Akbari, Pu Zhao, Weiwei Chen, David Kaeli, Yanzhi Wang

Late Breaking Results: Audio-Language-Action Models with Direct Speech Conditioning for Robotic Manipulation

DAC 2026

Enfu Nan, Pu Zhao, Yixiao Chen, Lin Zhao, Juyi Lin, Chen Wang, Weiwei Chen, Yanzhi Wang

LEAP: Learned Block-wise Evidence Retrieval for Long Audio-Video Perception

Preprint (arXiv:2609.39938)

Juyi Lin, Zhiqiang Lao, Jiali Cui, Lin Zhao, Pu Zhao, Dichang Zhang, Arman Akbari, Yu Qi, and others

ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation

NeurIPS 2025

Xiaomeng Yang*, Lei Lu*, Qihui Fan, Changdi Yang, Juyi Lin, Yanzhi Wang, Xuan Zhang, Shangqian Gao (* equal contribution)

Open-Source Acceleration of Stable-Diffusion.cpp Deployable on All Devices

Preprint

Jingxu Ng, Cheng Lv, Pu Zhao, Wei Niu, Juyi Lin, Minzhou Pan, Yun Liang, Yanzhi Wang

Reviewer & Program Committee

ICLR / ICANN / AAAI / WACV / EuroSys

• ICLR-27 Reviewer, NeurIPS-26 DiffuLM Workshop Reviewer, ICANN-26 Reviewer, AAAI-26/27 & WACV-27 LVR Workshop Program Committee, EuroSys'23 Shadow PC Reviewer
• Artifact Evaluation: MLSys'23, EuroSys'24/25, OSDI'24 & ATC'24
• Open Source Promotion Plan 2023