About Me

I am a first-year M.S. student in Computer Science & Technology at Xi'an Jiaotong University. I received my B.S. from Yanshan University, ranked 1st out of 191. My research focuses on large language models and multimodal / omni-modal foundation models. I have interned at ByteDance Seed and TikTok, working on high-quality data synthesis and ablation for omni-modal models as well as large-scale multimodal content understanding.

My research interests include:

• Multimodal & Omni-modal Foundation Models
• Continued Pre-training, Mid-training & Post-training
• Data Curation, Ablation Studies & Efficient Multimodal Reasoning

I am actively seeking research collaborations in LLMs and multimodal / omni-modal intelligence. I am deeply enthusiastic about ambitious, high-impact research—if our interests align, please feel free to reach out.

News

Recent Updates

Selected Publications

Bold = me. Links open the code / project page where available.

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

ACL 2026

Pengyu Li, Lingling Zhang, Zhitao Gao, Yanrui Wu, Yuxuan Dong, Huan Liu, Bifan Wei, Jun Liu.

Paper Code

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

Under Review

Pengyu Li, Zhitao Gao, Lingling Zhang, Muye Huang, Yuanming Li, Fangzhi Xu, Jun Liu.

Paper Code

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

ICANN 2024Oral

Pengyu Li, Chenhe Liu, Tengfei Li, Xinyu Wang, Shihui Zhang, Dongyang Yu.

Paper Code

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

COLM 2026

Zhitao Gao, Jie Ma, Xuhong Li, Pengyu Li, Ning Qu, Yaqiang Wu, Hui Liu, Jun Liu.

Paper Code

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

ACL 2026

Yuxuan Dong, Xinyu Zhang, Lingling Zhang, Han Lai, Pengyu Li, Bifan Wei, Yaqiang Wu, Jun Liu.

Paper Code

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

EMNLP 2026

Yanrui Wu, Lingling Zhang, Xinyu Zhang, Jiayu Chang, Pengyu Li, Xu Jiang, Jingtao Hu, Jun Liu.

Paper Code

GLaQ: Grounding Latent Queries in Visual Evidence for Multimodal Reasoning

Under Review

Zesheng Yang, Lingling Zhang, Xinyu Zhang, Cheng Zhang, Pengyu Li, Heng Wang, Lin Wu.

Paper

QUMem: Personalized Memory for Query-Conditioned User-State Inference in LLM Agents

Under Review

Heng Wang, Yifei Li, Lingling Zhang, Pengyu Li, Xinyu Che, Xinyu Zhang, Zesheng Yang.

Paper

Experience

Intern · ByteDance Seed

2026.02 – Present

Seed-Omni · Multimodal Data Synthesis & Ablation

Focused on high-quality multimodal data synthesis and ablation studies for Seed-Omni, building efficient, reusable pipelines for dialect speech, foundational S2T reasoning, long-form audio, and VS2T synthesis. Designed scalable production and quality-control workflows that improved key synthesis processes by up to 50×. Led controlled ablation studies for each newly introduced data type to isolate capability gains, regressions, and distribution conflicts. Systematically distilled insights from experimental behavior, evaluation results, and case analyses into reusable data recipes, quality criteria, and iteration guidelines for subsequent releases.

Intern · TikTok, ByteDance

2025.02 – 2025.09

EcoMLLM · Multimodal Content Understanding

Contributed EcoMLLM, a unified multimodal-LLM framework for large-scale short-video content understanding that replaced a stack of task-specific models. Designed a two-stage cascade—a multi-view VLM producing structured descriptions plus a compact LLM for fast downstream inference—and built Juggle Pipeline, a Ray + Merlin data system with 30+ operators for constructing large-scale instruction data. Combined training-free token pruning, FP8 quantization, and visual-token compression to increase QPS from 5.65 to 17.1 (>3×) without measurable downstream degradation. The system achieved 59.58% video-VQA accuracy and around 90% precision and recall on core tasks; RL (GRPO) with CoT cold-start and composite rewards improved US recall by 35% and instruction following to 100%.

Education

Honors & Awards

• First-class Scholarship — six consecutive semesters
• Siyuan Scholarship
• Provincial Merit Student (Hebei)
• CCPC (China Collegiate Programming Contest), Hebei — Silver Medal
• "Challenge Cup" National Business-Plan Competition — National Bronze
• National Energy-Saving & Emission-Reduction Competition — National 3rd Prize
• RAICOM Robot Developer Contest — National 3rd Prize
• National English Competition (NECCS) — National 2nd Prize
• National College Student Innovation & Entrepreneurship Training Program — National-level Project Lead
• 26th China Robotics and Artificial Intelligence Competition — Provincial 1st Prize
• 50+ national / provincial / university awards in total