Junxian Li

I'm a Ph.D. student at Shanghai Jiao Tong University, shanghai, China.

I'm now advised by associate prof. Yulun Zhang. I'm working on MLLMs, unified models and model reasoning. I've been advised by prof. Haojin Zhu in 2024.09-2026.03 (master). At Xi'an Jiaotong University I've worked on graph learning and data mining. I've worked as a remote research intern at DaRL Group with assistant prof. Hua Wei for one year (also co-advised by assistant prof. Bin Shi). I've also collaborated with Dr. Sichao Li, Di Zhang and Jiatong Li.

Email:  ljx201806 [at] gmail [dot] com  /  lijunxian0531 [at] sjtu [dot] edu [dot] cn  /  ljx201806 [at] stu [dot] xjtu [dot] edu [dot] cn

Email  /  CV  /  Bio  /  Scholar  /  Github

Research /

I'm interested in MLLMs & UMMs, LLM security, and data mining. Representative papers are highlighted. Feel free to drop me an email for possible collaboration, especially on MLLM reasoning and MLLM security!

This paper proposes G²TR, a generation-guided token reduction method for UMMs.

This paper proposes δ-mem, a memory mechanism that keeps a compact and dynamically updated memory alongside a frozen full-attention backbone.

This paper proposes PermuQuant, a method for low-bit quantization based on channel ordering.

We introduce SlowBA, an efficiency backdoor on VLM-based GUI agents. Experiments and real-world studies show the real vulnerabilities.

We introduce S²-bench, a benchmark for open-ended molecule generation, benefiting molecule discovery and medicine design.

We introduce AttnDiff, an attention-based LLM watermarking technique.

We introduce FaithEval and FaithAct, a unified framework that evaluates and enforces step-level perceptual faithfulness in multimodal reasoning, significantly improving grounding and stability without sacrificing task performance.

We introduce StreamMeCo, a training-free framework for compressing long-term agent memory, enabling efficient and accurate streaming video understanding.

This paper reveals a novel and realistic vulnerability: the first multi-target backdoor attack on VLM-based visual grounding.

We propose S²-MLLM, an efficient framework that enhances spatial reasoning in MLLMs through implicit spatial reasoning.

we introduce Biology-Instructions, the first large-scale instruction-tuning dataset for multi-omics biological sequences, including DNA, RNA, proteins, and multi-molecules.

We introduce Critic-V, a novel framework inspired by the Actor-Critic paradigm to boost the reasoning capability of VLMs.

We propose ChemVLM, an open-source multimodal large language model for chemical tasks.

Uncertainty quantification (like dropout and evidential deep learning) can be used for better sensor deployment on spatio-temporal traffic prediction.

Using DQNs to solve traffic signal control problems under missing data scenario.

We introduce LoRA-Chem, a modular LoRA-based framework for accurate and scalable multitask prediction across diverse organic reactions.

We propose LT-OPD, an on-policy self-distillation framework for extreme visual token reduction, enabling MLLMs to retain strong capabilities under ultra-low visual-token budgets.

We introduce AI4Service and Alpha-Service, a unified framework for proactive assistance through AI glasses, combining real-time perception, task orchestration, tool use, and long-term memory to anticipate user needs and deliver personalized services.

Education

Honors&Awards

Services&Academic Experiences

Hobbies

Feel free to clone this website's source code. Do not scrape the HTML from this page itself, as it includes analytics tags that you do not want on your own website — use the github code instead. Also, consider using Leonid Keselman's Jekyll fork of this page.