Pengxiang Li (李鹏翔)
Ph.D. Student @ PolyU
I am a Ph.D. student at The Hong Kong Polytechnic University, advised by Prof. Hongxia Yang. Previously, I received my MSc in Artificial Intelligence from Dalian University of Technology, advised by Prof. Huchuan Lu.
My research focuses on how generative models learn, reason, and act efficiently. I study language model training and inference, with a particular interest in diffusion language models and parallel decoding. I also work on multimodal agents for GUI interaction and autonomous driving, connecting visual understanding with reasoning and action generation. My earlier work explored controllable video generation and the evaluation of vision-language models in challenging driving scenes.
- May 2026 New preprint: ReflectDrive-2 uses reinforcement learning to align trajectory generation and self-editing in a discrete diffusion planner.
- Feb 2026 New preprint: NAP investigates why diffusion language models struggle with parallel decoding and how training data can help.
- EACL 2026 InfiGUIAgent is published in the EACL 2026 main conference proceedings.
- ICLR 2026×2 Two papers are accepted at ICLR 2026: Discrete Diffusion for Reflective Vision-Language-Action Models in Autonomous Driving and Diffusion Language Models Know the Answer Before Decoding !
- AAAI 2026 One paper ( InfiGUI-G1: Advancing GUI Grounding with Adaptive Exploration Policy Optimization ) is accepted to AAAI 2026 (Oral)!
- ICRA 2026 One paper ( TransDiffuser: End-to-end trajectory generation with decorrelated multi-modal representation for autonomous driving ) is accepted to ICRA 2026!
- NeurIPS 2025×3 Three papers have been accepted at NeurIPS 2025: Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking , The Curse of Depth in Large Language Models , and GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling
- ICLR 2025 One paper ( Mix-LN: Unleashing the Power of Deeper Layers by Combining Pre-LN and Post-LN ) is accepted to ICLR 2025!
- ACL 2025 One paper ( Outlier-weighed Layerwise Sampling for LLM Fine-tuning ) is accepted to ACL 2025 Findings!
- WACV 2025×2 Two papers are accepted to WACV 2025 on VLM evaluation ( Automated Evaluation of Large Vision-Language Models on Self-driving Corner Cases ) and Controllable Generation ( TrackDiffusion )!
- Efficient Language Models: Training dynamics, normalization, depth scaling, and efficient fine-tuning.
- Diffusion Language Models: Parallel decoding, early answer convergence, and controllable generation.
- Multimodal Agents: GUI grounding and reasoning; vision-language-action models, trajectory generation, and grounded risk understanding for autonomous driving.