Hello! I'm Shicheng Xu (徐士成), a Post-training Researcher at DeepSeek, focusing on reinforcement learning and agentic post-training for large language models.
I received my Ph.D. from the Institute of Computing Technology, Chinese Academy of Sciences (State Key Laboratory of AI Safety), under the supervision of Prof. Liang Pang and Prof. Xueqi Cheng, and my B.S. from Harbin Institute of Technology. Before joining DeepSeek, I was an intern at ByteDance Seed (TopSeed Talent Program), where I worked for RL scaling in the post-training of Seed 1.8, Seed 2.0 and Seed 2.1.
My research interests include Reinforcement Learning (RL), Agent, and Retrieval-Augmented Generation (RAG).
🔥 News
- 2026.07: I joined DeepSeek as a Post-training Researcher, working on RL and agentic post-training.
- 2026.07: I received my Ph.D. degree from ICT, CAS. Grateful to my advisors and collaborators!
- 2026.06: 🚀 Seed 2.1 is released (Seed2.1 Model Card: Agentic Intelligence for Productivity). I worked for RL scaling in its post-training as a core contributor.
- 2026.04: 🎉 Our paper DxDirector: An Agentic Large Language Model Driving the Full-process Clinical Diagnosis is accepted by Nature Communications.
- 2026.02: 🚀 Seed 2.0 is released (Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity). I worked for RL scaling in its post-training as a core contributor.
- 2025.12: 🚀 Seed 1.8 is released (Seed1.8 Model Card: Towards Generalized Real-World Agency). I worked for RL scaling in its post-training as a core contributor.
- 2025.12: 🎉 Our paper RLKD: Distilling LLMs' Reasoning via Reinforcement Learning is accepted by AAAI 2026.
- 2025.11: I joined ByteDance Seed as a TopSeed intern, working on RL scaling for Seed foundation models.
🚀 Foundation Model Technical Reports
Frontier foundation models I contributed to as a core member of the RL scaling / post-training team at ByteDance Seed.
Seed2.1 Model Card: Agentic Intelligence for Productivity Technical Report
Worked for RL scaling algorithms in post-training (Core Contributor).
Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity Technical Report
Worked for RL scaling in post-training (Core Contributor).
Seed1.8 Model Card: Towards Generalized Real-World Agency Technical Report
Worked for RL scaling in post-training (Core Contributor).
📝 Publications
First-author publications. Full list on Google Scholar.
DxDirector: An Agentic Large Language Model Driving the Full-process Clinical Diagnosis Nature Communications
A 7B clinical agentic LLM that outperforms much more expensive models (DeepSeek, o3-mini, o1, Gemini-2.0-flash) on dynamic, complex, end-to-end full-process clinical reasoning, while interacting with real-world medical workflows efficiently and robustly.
A Theory for Token-Level Harmonization in Retrieval-Augmented Generation ICLR 2025
The first theoretical framework for the fusion between LLMs' parametric knowledge and retrieved external knowledge, explaining the underlying mechanism and enabling a robust collaborative generation framework.
Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models ICLR 2025
Identifies and fixes the hidden alignment gap between visual and textual modalities in LVLMs, improving the robustness of multimodal agents against harmful visual inputs.
Reshapes the interaction paradigm between LLMs and external tools: keeps the LLM's deep reasoning ability while fully exploiting external information, with robustness to noisy environments.
An unsupervised training-data construction method that aligns LLMs with knowledge fed back from external environments, achieving efficient and robust fusion of internal and external knowledge.
List-aware Reranking-Truncation Joint Model for Search and Retrieval-augmented Generation WWW 2024 Oral
A generative joint reranking-truncation model that provides high-quality retrieval lists for RAG and keeps RAG robust to retrieval results.
Invisible Relevance Bias: Text-Image Retrieval Models Prefer AI-Generated Images SIGIR 2024 Oral
Reveals the hidden fairness impact of AI-generated images on cross-modal retrieval models.
BERM: Training the Balanced and Extractable Representation for Matching to Improve Generalization Ability of Dense Retrieval ACL 2023 Oral
Improves multi-task, multi-domain generalization of dual-encoder retrieval models by explicitly modeling relevance-scoring behavior during contrastive learning.
Match-Prompt: Improving Multi-task Generalization Ability for Neural Text Matching via Prompt Learning CIKM 2022 Oral
Improves multi-task, multi-domain generalization of text matching models via prompt learning in continuous vector space.
NIR-Prompt: A Multi-task Generalized Neural Information Retrieval Training Framework TOIS
A full-pipeline training framework for generalizable neural information retrieval, from recall to ranking.
RLKD: Distilling LLMs' Reasoning via Reinforcement Learning AAAI 2026
Effective distillation of LLMs' reasoning ability via reinforcement learning.
📚 Tutorials
- Bias and Unfairness in Information Retrieval Systems: New Challenges in the LLM Era (KDD 2024, WSDM 2025, SIGIR-AP 2025) [Website] [Survey] [GitHub]
🎖 Honors and Awards
- CAS President Award (中国科学院院长奖)
- Hua Luogeng Scholarship (华罗庚奖学金)
- Beijing Outstanding Graduate (北京市优秀毕业生)
- UCAS Outstanding Graduate (中国科学院大学优秀毕业生)
- ICT Director's Special Award (中国科学院计算技术研究所所长特别奖)
- National Scholarship (国家奖学金)
- UCAS First-class Academic Scholarship (中国科学院大学一等学业奖学金)
- WWW 2024 Distinguished Speaker
- HIT Outstanding Graduate (哈尔滨工业大学优秀毕业生)
📖 Educations
- 2021.09 - 2026.07, Institute of Computing Technology, Chinese Academy of Sciences
- Ph.D., State Key Laboratory of AI Safety
- 2017.09 - 2021.07, Harbin Institute of Technology
- B.S., School of Computer Science and Technology
💻 Experience
- 2026.07 - Present, DeepSeek, Post-training Researcher
- Post-training algorithms for large language models: reinforcement learning and agentic training.
- 2025.11 - 2026.06, ByteDance Seed , Intern, TopSeed Talent Program
- RL scaling for the post-training of Seed 1.8, Seed 2.0 and Seed 2.1.