CV
Education
- MPhil in Computer Science, Peking University, 2025 – 2028
- School of Computer Science / Institute of Computational Linguistics
- Advisor: Prof. Yunfang Wu
- Research: Natural Language Processing, Large Language Models, Reinforcement Learning
- B.S. in Computer Science, Beijing Institute of Technology, 2021 – 2025
- Xu Teli Honors Program
GPA: 3.8 / 4.0 Rank: 1 / 68
Publications
Zhaoxi Zhang, Sanwoo Lee, Zhixiang Wang, Yunfang Wu. 3DM: Distill, Dynamic Drop, and Merge for Debiasing Multi-modal Large Language Models. ACL 2025.
Yihan Lin*, Zhaoxi Zhang*, Taoyi Wang, Yuguo Chen, Rong Zhao. CSVO: Complementary-Pathway Spatial-Enhanced Visual Odometry for Extreme Environments with Brain-inspired Vision Sensors. IROS 2025. (Co-first author)
Yanzhi Zhang*, Zhaoxi Zhang*, Haoxiang Guan, Yilin Cheng, Yitong Duan, Chen Wang, Yue Wang, Yunfang Wu, Shuxin Zheng, Jiyan He. No Free Lunch: Rethinking Internal Feedback for LLM Reasoning. (Co-first author, under review)
Zhaoxi Zhang, Yitong Duan, Yunfang Wu. One Tool Is Enough: Reinforcement Learning of LLM Agents for Repository-level Code Navigation. (Under review)
(For a full list, see Google Scholar.)
Research & Internship Experience
Reinforcement Learning for LLM Agents, Baidu Research Jan 2026 – Present - Researching end-to-end RL training and evaluation of LLM agents.
- Agentic Learning with Skills: Designed a two-agent system (Skill Generator + Solver). The Solver is trained with GRPO; the Generator receives reward based on the performance gain the skills provide and is updated with REINFORCE. Achieved improvements on multiple single-hop and multi-hop QA benchmarks. Targeting NeurIPS / EMNLP.
- Repository-level Code Generation Benchmark: Designing a cross-language benchmark for evaluating LLMs’ repo-level code generation capability. Successfully constructed 400 samples and scaling up. Targeting NeurIPS / EMNLP.
- Mixture of Experts as Multi-Agents: Exploring sparse MoE architectures to design novel MARL algorithms where each agent role activates distinct experts, aiming to resolve capability conflicts in multi-agent systems.
Reinforcement Learning for LLM Agents, Zhongguancun Institute of Artificial Intelligence (ZIAI) Apr 2025 – Jan 2026 - Studied whether RL teaches LLMs genuinely new skills in the agentic setting, where tools are newly designed to avoid data leakage.
- Zoom-R1: Proposed a visual reasoning method where the model invokes a zoom tool to insert magnified sub-images into its chain-of-thought, enabling fine-grained spatial reasoning over images.
- CodeBase QA (One Tool Is Enough): Redefined SWE-Bench as a code localization task. Trained an agent via multi-turn tool-call RL using the VeRL framework, allowing the model to browse source code during rollouts. Submitted in January 2026.
Model Merging for Debiasing Multimodal LLMs, Peking University Nov 2024 – Feb 2025 - Investigated debiasing multimodal LLMs via model merging. Used knowledge distillation to unify heterogeneous model architectures, then designed a novel merging strategy that preserves capability-enhancing parameters while discarding bias-inducing ones. Published at ACL 2025.
Multimodal Visual Odometry, Tsinghua University Nov 2023 – Mar 2025 - Addressed the failure of RGB-only VO under extreme conditions (HDR, high speed) using the Tianmouc neuromorphic sensor, which outputs Time-Difference (TD) and Space-Difference (SD) streams. Designed an asynchronous multimodal fusion network combining SD and RGB channels. Achieved state-of-the-art results on standard benchmarks and real Tianmouc recordings. Published at IROS 2025.
Awards & Honors
- Excellence Scholarship (First Prize), Beijing Institute of Technology
- National English Competition for College Students — First Prize, Beijing
- “FLTRP·ETIC Cup” English Speaking Contest — First Prize, Beijing; 1st Place, Beijing Institute of Technology