DU Yiming

Huawei Hong Kong Research Institute

LegoX Team

Hong Kong, China

Hi, I am Yiming (Elvin) DU, a researcher at Huawei Hong Kong Research Institute, a member of the LegoX team, and a core contributor to LEGO-RL. My research focuses on coding agents, agentic reinforcement learning, and large language model agents. Recently, I have been working on long-horizon agent RL in real-world software engineering environments, harness-native training, and consistency between training and inference.

I received my Ph.D. from The Chinese University of Hong Kong, where I was supervised by Prof. Kam-Fai Wong in the Department of SEEM and the MoE Key Laboratory of High Confidence Software Technologies. From January to June 2025, I visited EdinburghNLP, working with Prof. Jeff Z. Pan. Previously, I interned at Huawei Noah’s Ark Lab and SenseTime Research, and conducted research at Imperial College London and City University of Hong Kong. I received my M.Eng. and B.Eng. in Computer Science from Southeast University.

My broader goal is to build agents that learn from interaction, retain useful experience, and reliably solve complex tasks over time. My work has appeared in venues including ICLR, AAAI, EMNLP, and ACM Computing Surveys.

Research Interests

Coding Agents & Agentic Reinforcement Learning. Long-horizon learning in real software engineering environments, harness-native training, and training–inference consistency. I study how to train reliable coding agents within the tools and execution workflows they use at deployment.
Representative works: LEGO-RL, SWE-Lego.
Memory & Reasoning for LLM Agents. Long-term and multimodal memory, temporal reasoning, and retrieval over extended interactions, connecting my earlier work on personal assistants with agents that learn from experience.
Representative works: MEMORY-T1 (ICLR 2026), MemGuide (AAAI 2026), MemLens, Multimodal Retrieval Heads, PerLTQA (Best Paper, SIGHAN 2024), Rethinking Memory in AI.
Data & Learning for Interactive Agents. High-quality trajectory and dialogue data, robust supervised fine-tuning, and personalized multi-turn interaction, spanning software engineering agents and conversational assistants.
Representative works: SWE-Lego, ReSURE (EMNLP 2025), Flexibly Utilize Memory (EMNLP 2025), Sticker Retrieval (AAAI 2025).

Featured Project

LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents brings reinforcement learning to native coding-agent harnesses, with an emphasis on faithful optimization, reliable execution, and observable training.

Paper · GitHub · Hugging Face · Project / Docs

Honors & Awards

2024 — Overseas Research Attachment Programme (ORAP) Award, CUHK.
2024 — Best Paper Award, ACL-SIGHAN 2024 (for PerLTQA).
2020 — 1st Place, Jiangsu Postgraduate Mathematical Modeling Innovation Practice Competition.
2018 – 2020 — Excellent Scholarships: Shenzhen Stock Exchange, Southeast University, Huawei, Keyuan Automation.
2018 — 1st Place, Chinese Undergraduate Computer Design Contest; Meritorious Winner, Mathematical Contest in Modeling (MCM).

Academic Services

Reviewer: ACL, EMNLP, IJCAI, Knowledge-Based Systems (KBS), etc.
Patent: A dialogue memory storage, extraction and electronical device — Baojun Wang, Yiming Du, Kam-Fai Wong (Huawei, 2024).

news

Sep 24, 2026	I have received my Ph.D. from The Chinese University of Hong Kong, supervised by Prof. Kam-Fai Wong. I am now a researcher at Huawei Hong Kong Research Institute, a member of the LegoX team, and a core contributor to LEGO-RL.
Aug 18, 2026	We introduce LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents. Read the paper and explore the code, models, and documentation.
May 26, 2026	New collaborative work: Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models. Read the paper.
May 14, 2026	📄 New paper MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models released on arXiv.
Jan 22, 2026	🎉 Our paper MEMORY-T1: Reinforcement Learning for Temporal Reasoning in Multi-Session Agents is accepted to ICLR 2026.
Jan 04, 2026	Our paper SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving is available on arXiv. We study supervised fine-tuning for software engineering agents.

selected publications

LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents

Yiming Du, Yuxin Jiang, Tao Yuan, and 9 more authors

arXiv preprint arXiv:2608.17393. Hugging Face , 2026

arXiv Bib Code Website

@article{du2026legorl,
  title = {{LEGO-RL}: Harness-Native Reinforcement Learning for Coding Agents},
  author = {Du, Yiming and Jiang, Yuxin and Yuan, Tao and Dai, Jianbo and Wang, Shaowei and Chen, Jierun and Tao, Chaofan and Yu, Xianzhi and Shang, Lifeng and Wong, Kam-Fai and Li, Xiaohui and Bai, Haoli},
  journal = {arXiv preprint arXiv:2608.17393},
  year = {2026},
}

Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models

Aaron Branson Cigres Li, Zhaowei Wang, Yu Zhao, and 9 more authors

arXiv preprint arXiv:2605.27243, 2026

arXiv Bib

@article{li2026retrievalheads,
  title = {Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models},
  author = {Li, Aaron Branson Cigres and Wang, Zhaowei and Zhao, Yu and Du, Yiming and Li, Haobo and Ren, Xiyu and Wong, Ginny and See, Simon and Luo, Lishu and Duan, Haodong and Minervini, Pasquale and Song, Yangqiu},
  journal = {arXiv preprint arXiv:2605.27243},
  year = {2026},
}

SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving

Chaofan Tao, Jierun Chen, Yuxin Jiang, and 11 more authors

arXiv preprint arXiv:2601.01426, 2026

arXiv Bib

@article{tao2026swelego,
  title = {{SWE-Lego}: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving},
  author = {Tao, Chaofan and Chen, Jierun and Jiang, Yuxin and Kou, Kaiqi and Wang, Shaowei and Wang, Ruoyu and Li, Xiaohui and Yang, Sidi and Du, Yiming and Dai, Jianbo and Mao, Zhiming and Wang, Xinyu and Shang, Lifeng and Bai, Haoli},
  journal = {arXiv preprint arXiv:2601.01426},
  year = {2026},
}

MEMORY-T1: Reinforcement Learning for Temporal Reasoning in Multi-Session Agents

Yiming Du, Baojun Wang, Yifan Xiang, and 11 more authors

In International Conference on Learning Representations (ICLR), 2026

Bib

@inproceedings{du2026memoryt1,
  title = {MEMORY-T1: Reinforcement Learning for Temporal Reasoning in Multi-Session Agents},
  author = {Du, Yiming and Wang, Baojun and Xiang, Yifan and Wang, Zhaowei and Huang, Wenyu and Xue, Boyang and Liang, Bin and Zeng, Xingshan and Bai, Haoli and Mi, Fei and Shang, Lifeng and Pan, Jeff Z and Jiang, Yuxin and Wong, Kam-Fai},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2026},
}

MemGuide: Intent-Driven Memory Selection for Goal-Oriented Multi-Session LLM Agents

Yiming Du, Bingbing Wang, Yang He, and 7 more authors

In Proceedings of the AAAI Conference on Artificial Intelligence, 2026

Accepted, Oral. Co-first author with Bingbing Wang.

Bib

@inproceedings{du2026memguide,
  title = {MemGuide: Intent-Driven Memory Selection for Goal-Oriented Multi-Session LLM Agents},
  author = {Du, Yiming and Wang, Bingbing and He, Yang and Liang, Bin and Wang, Baojun and Li, Zhongyang and Gui, Lin and Pan, Jeff Z and Xu, Ruifeng and Wong, Kam-Fai},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year = {2026},
  note = {Accepted, Oral. Co-first author with Bingbing Wang.},
}

ReSURE: Regularizing Supervision Unreliability for Multi-turn Dialogue Fine-tuning

Yiming Du, Yifan Xiang, Bin Liang, and 3 more authors

In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025

Bib

@inproceedings{du2025resure,
  title = {ReSURE: Regularizing Supervision Unreliability for Multi-turn Dialogue Fine-tuning},
  author = {Du, Yiming and Xiang, Yifan and Liang, Bin and Lin, Dahua and Wong, Kam-Fai and Tan, Fei},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year = {2025},
}

Flexibly Utilize Memory for Long-Term Conversation via a Fragment-then-Compose Framework

Cai Ke, Yiming Du, Bin Liang, and 8 more authors

In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025

Accepted, Oral.

Bib

@inproceedings{ke2025flexibly,
  title = {Flexibly Utilize Memory for Long-Term Conversation via a Fragment-then-Compose Framework},
  author = {Ke, Cai and Du, Yiming and Liang, Bin and Yifan, Xiang and Gui, Lin and Li, Zhongyang and Wang, Baojun and Yu, Yue and Wang, Hui and Wong, Kam-Fai and Xu, Ruifeng},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year = {2025},
  note = {Accepted, Oral.},
}

Rethinking Memory in AI: Taxonomy, Operations, Topics, and Future Directions

Yiming Du, Wenyu Huang, Danna Zheng, and 5 more authors

arXiv preprint arXiv:2505.00675, 2025

Bib

@article{du2025rethinking,
  title = {Rethinking Memory in AI: Taxonomy, Operations, Topics, and Future Directions},
  author = {Du, Yiming and Huang, Wenyu and Zheng, Danna and Wang, Zhaowei and Montella, Sebastien and Lapata, Mirella and Wong, Kam-Fai and Pan, Jeff Z},
  journal = {arXiv preprint arXiv:2505.00675},
  year = {2025},
}

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models

Xiyu Ren, Zhaowei Wang, Yiming Du, and 11 more authors

arXiv preprint arXiv:2605.14906, 2026

arXiv Bib Code

@article{ren2026memlens,
  title = {MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models},
  author = {Ren, Xiyu and Wang, Zhaowei and Du, Yiming and Xie, Zhongwei and Liu, Chi and Yang, Xinlin and Feng, Haoyue and Pan, Wenjun and Zheng, Tianshi and Xu, Baixuan and Li, Zhengnan and Song, Yangqiu and Wong, Ginny and See, Simon},
  journal = {arXiv preprint arXiv:2605.14906},
  year = {2026},
}

A New Formula for Sticker Retrieval: Reply with Stickers in Multi-Modal and Multi-Session Conversation

Bingbing Wang, Yiming Du, Bin Liang, and 4 more authors

In Proceedings of the AAAI Conference on Artificial Intelligence, 2025

Co-first author.

Bib

@inproceedings{wang2025new,
  title = {A New Formula for Sticker Retrieval: Reply with Stickers in Multi-Modal and Multi-Session Conversation},
  author = {Wang, Bingbing and Du, Yiming and Liang, Bin and Bai, Zhixin and Wang, Baojun and Xu, Ruifeng and Wong, Kam-Fai},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume = {39},
  number = {24},
  pages = {25327--25335},
  year = {2025},
  note = {Co-first author.},
}

A Comprehensive Study of Multilingual Confidence Estimation on Large Language Models

Boyang Xue, Hongru Wang, Rui Wang, and 4 more authors

In Findings of the Association for Computational Linguistics: ACL 2025, 2025

Bib

@inproceedings{xue2025multilingual,
  title = {A Comprehensive Study of Multilingual Confidence Estimation on Large Language Models},
  author = {Xue, Boyang and Wang, Hongru and Wang, Rui and Wang, Sheng and Wang, Zezhong and Du, Yiming and Wong, Kam-Fai},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  year = {2025},
}

PerLTQA: A Personal Long-Term Memory Dataset for Memory Classification, Retrieval, and Fusion in Question Answering

Yiming Du, Hongru Wang, Zhengyi Zhao, and 5 more authors

In Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10), 2024

Best Paper Award.

Bib

@inproceedings{du2024perltqa,
  title = {PerLTQA: A Personal Long-Term Memory Dataset for Memory Classification, Retrieval, and Fusion in Question Answering},
  author = {Du, Yiming and Wang, Hongru and Zhao, Zhengyi and Liang, Bin and Wang, Baojun and Zhong, Wanjun and Wang, Zezhong and Wong, Kam-Fai},
  booktitle = {Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10)},
  pages = {152--164},
  year = {2024},
  note = {Best Paper Award.},
}