Hi there, I'm Pengyu Li (李鹏宇) 👋

I am a first-year M.S. student in Computer Science & Technology at Xi'an Jiaotong University. My research centers on large language models and multimodal / omni-modal foundation models, with a focus on the full post-training pipeline.

I have interned at ByteDance Seed, TikTok, and iFLYTEK, working on omni-modal (speech) foundation models, large-scale multimodal content understanding, and medical LLMs.

• 🔭 Currently working on the speech modality of an omni-modal foundation model @ ByteDance Seed
• 🌱 Interested in unifying perception, reasoning, and generation across modalities — efficiently
• 💬 Happy to chat about multimodal LLMs, omni-modal training, and RL post-training
• 📫 Feel free to email me for any form of academic cooperation!

• Multimodal & Omni-modal LLMs — unifying text, vision, and speech into a single foundation model
• LLM Continued Pre-training · Mid-training · Post-training — data recipes, task composition & interference
• Efficient & Unified Multimodal Reasoning — token pruning, distillation, RL (GRPO) for reasoning

^⭐ = first author  ·  full list on Google Scholar

• ^⭐ AGT^AO: Robust and Stabilized LLM Unlearning via Adversarial Gating Training with Adaptive Orthogonality
   

• ^⭐ Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning
   

• ^⭐ EMDFNet: Efficient Multi-scale and Diverse Feature Network for Traffic Sign Detection
   

• PhysPRM: A Generative Process Reward Model with Fine-grained Diagnosis for Physics Problem Solving
   

• AERO: Autonomous Evolutionary Reasoning Optimization via Endogenous Dual-Loop Feedback
   

• LogicGraph: Benchmarking Multi-Path Logical Reasoning via Neuro-Symbolic Generation and Verification
   

_{"Long-term correctness over short-term convenience."}