Junli Gong

I am an MS student in Computer Science at Northeastern University (Khoury College, Seattle), graduating in December 2026. I received my BSc (Honours) in Computer Science and Technology from Hong Kong Baptist University.

At Khoury I work with Prof. Yifan Hu on speculative expert pre-dispatch for Mixture-of-Experts inference. Since 2024 I have worked with Prof. Weifeng Su at Beijing Normal-Hong Kong Baptist University (BNBU) on multimodal perception and embodied agents.

In summer 2025 I was a research intern at Pazhou Lab / South China University of Technology, advised by Prof. Lu Lu, writing Tensor Core GEMM kernels for Transformer workloads and benchmarking them against cuBLAS. In summer 2026 I was an AI Infra intern at iFLYTEK, adapting speech models to heterogeneous domestic AI accelerators. Earlier, I co-first-authored a WASA 2025 paper on clustered federated learning with Prof. Jianxiong Guo at Beijing Normal University, Zhuhai, and worked with Dr. Nian Xia at Nanjing Normal University on ML-based malicious traffic detection (AsiaJCIS 2024).

I will present SegWorld (oral) and PEAM (poster) at EMNLP 2026 in November.

Research
Interest

I work on efficient, scalable systems for multimodal and mixture-of-experts models — LLM inference, communication-aware distributed inference, and hardware-aware acceleration.

Building multimodal perception and embodied-agent models taught me what these workloads look like; the systems work is about making them run fast without changing what they compute.

LLM Inference Systems GPU Kernels Multimodal Perception Embodied Agents

LLM Inference Systems

Serving large models is often bound by communication rather than compute. I study how routing and scheduling decisions can be predicted early enough to overlap communication with computation, without changing what the model outputs.

Speculative expert pre-dispatch for MoE inference (in progress)

GPU Kernels & Heterogeneous Accelerators

What a hand-written kernel can still win against a vendor library, and what breaks when a model moves from one accelerator to another.

Tensor Core HGEMM vs. cuBLAS (Pazhou Lab)·Cross-accelerator model adaptation (iFLYTEK)

Multimodal Perception

Fusing what different sensors see — thermal and visible, multiple medical modalities — into something a downstream model can act on, and judging fusion quality the way people do.

FusionProxy·FuScore·Fuse4Seg

Embodied Agents

Whether a model can reason from an intent to the part of the scene that affords it, and how an agent's experience becomes part of its parameters rather than a buffer it looks up.

SegWorld·PEAM

Publications

Peer-reviewed

Can Segmentation Models Understand the World? Proactive Affordance Reasoning via Vision-Grounded Chain-of-Thought

Yuchen Guo, Junli Gong, Hongmin Cai, Yiu-ming Cheung, Weifeng Su

Empirical Methods in Natural Language Processing (EMNLP), 2026 Main Oral Paper

PEAM: Parametric Embodied Agent Memory for Continual Skill Learning via Internalization of Experience in Minecraft

Yuchen Guo*, Junli Gong*, Weicheng Wang, Hongmin Cai, Yiu-ming Cheung, Weifeng Su

Empirical Methods in Natural Language Processing (EMNLP), 2026 Findings Poster Paper

Adding Thermal Awareness to Visual Systems in Real-Time via Distilled Diffusion Models

Yuchen Guo, Junli Gong, Wenjun Dong, Yiu-ming Cheung, Weifeng Su

CVPR Workshop on World Models Meet Active Sensing (WMAS), 2026 Poster Paper

Enhancing Federated Learning in IoT Through Dynamic Spectral Clustering

Changyu Zheng*, Junli Gong*, Jianxiong Guo, Zhiqing Tang, Tian Wang

Wireless AI Computing Systems and Applications (WASA), 2025 Oral Paper

An Effective Feature Selection Algorithm for Machine Learning-based Malicious Traffic Detection

Chao Fei, Nian Xia, Pang-Wei Tsai, Yang Lu, Xiaonan Pan, Junli Gong

Asia Joint Conference on Information Security (AsiaJCIS), 2024 Paper

Preprints & under review

Bringing Multimodal Large Language Models to Infrared-Visible Image Fusion Quality Assessment

Yuchen Guo*, Junli Gong*, Yao Lu, Xintong Xu, Yiu-ming Cheung, Weifeng Su

Preprint, 2026 Paper

LumiVideo: An Intelligent Agentic System for Video Color Grading

Yuchen Guo, Junli Gong, Yiu-ming Cheung, Weifeng Su

Preprint, 2026 Project Paper

Fuse4Seg: Image Fusion for Multi-Modal Medical Segmentation via Bi-level Optimization

Yuchen Guo, Junli Gong, Yiu-ming Cheung, Weifeng Su

Preprint, 2024 Paper

* denotes equal contribution.

The purpose of computing is insight, not numbers. — Richard Hamming