Qianli Ma

I am a Ph.D. student at Shanghai Jiao Tong University, advised by Prof. Zhipeng Zhang . I completed my master's degree in Computer Science at Shanghai Jiao Tong University advised by Prof. Li Niu and Prof. Linfeng Zhang and my bachelor's degree in Instrument Science and Control Technology at Southeast University.

I'm interested in LLM posttraining and multimodal learning. I am also interested in multi-agent systems.

I am actively seeking collaborations on exploring reasoning ability of LLMs and multi-agent systems, please feel free to contact me!!

Google Scholar  /  Semantic Scholar  /  GitHub  /  Twitter  /  Email

^* denotes equal contribution, ^† denotes corresponding author, some are highlighted.

🤗 HF Paper

To address the questions "How to create a webpage from an academic paper?" and "How to evaluate the project webpage?", we propose AutoPage and PageBench. AutoPage transforms academic papers into polished, published-ready project webpages through a human-in-the-loop multi-agent pipeline, while PageBench provides automatic evaluation across content quality and visual design quality dimensions.

This paper proposes a new finetuning method for diffusion models, which decouples the diffusion process into multiple denoising tasks and then merges them. We show that this method can effectively finetune diffusion models for various tasks, including text-to-image generation, unconditional image generation.

This paper proposes an efficient diffusion model for low light enhancement, which can be applied to various low light enhancement tasks.

This paper proposes a method to mitigate safety-utility conflicts in model merging for LLMs, which can be applied to various safety-utility tasks.

🤗 HF Paper

This paper proposes a comprehensive and versatile benchmark suite for video generative models.

This paper proposes a token pruning method for stable diffusion, which can accelerate the generation process by 9 times.

We present RAPO++, a cross-stage prompt optimization framework that unifies training-data--aligned refinement, test-time iterative scaling, and large language model (LLM) fine-tuning to substantially improve T2V generation without modifying the underlying generative backbone.

We present a comprehensive survey of general end-to-end autonomous driving. This survey collects and organizes key papers in General End-to-End Autonomous Driving, classifying them into Conventional (e.g., UniAD), VLM-centric (e.g., DriveLM), and Hybrid (e.g., Senna) approaches. In addition, this survey curates both Normal and Vision-Language datasets relevant to General End-to-End Autonomous Driving. Based on this taxonomy and dataset collection, our analysis further outlines the main research branches and emerging trends that are shaping the field.

• Teaching Assistant: AI7002 AI Ethics, SJTU, 2025 Autumn
• Teaching Assistant: AI2805 Digital Signal and Image Processing, SJTU, 2026 Spring

• Conference Reviewer: ICML, ICCV, CVPR, ECCV, ARR, NeurIPS
• Award: Top 3 in IROS 2025 RoboSense Challenge
• Award: First-Class Graduate Academic Scholarship in 2023, 2024
• Award: Freshman Graduate Academic Scholarship in 2022
• Award: Undergraduate Course Scholarship in 2020, 2021

Derived from Jon Barron's website and Rui Ye's website.