Yirong Zeng (曾屹荣)

Ph.D. Candidate · Agentic RL / Instruction Following
Open to internships & campus hiring · Let's talk

Yirong Zeng (曾屹荣)

I am a direct-PhD student at the Harbin Institute of Technology main campus, in the Social Computing and Information Retrieval (SCIR) Lab (advisor: Prof. Ting Liu). My research focuses on Agentic RL post-training and instruction following for large language models. I have published 9 first-author (incl. co-first) papers at NeurIPS, ICLR, ACL, EMNLP, NAACL, COLING and IPM — 4 CCF-A, 4 CCF-B, 1 SCI-Q1 — with 6 more CCF-A papers under review. I have completed a foundation-model reinforcement learning post-training internship at a major tech company.

Research Interests

Research Interests
Agentic Reinforcement Learning 2024 - Present
  • Acquiring environment and human feedback through tool calls and continuously updating model parameters to raise the performance ceiling — a mainstream post-training route for large models.
  • Research coverage: stability and ceiling of long-horizon training (NeurIPS 2026, EMNLP 2025×2, ICLR 2027 under review), adaptive fast/slow thinking (ICLR 2026), tool environments and data synthesis (NeurIPS 2026, ICLR 2027 under review), and lightweight Harness with In-Harness RL (AAAI 2027 under review).
  • Recent exploration: native Harness RSI implementation and research.
Instruction Following 2025.06 - Present
  • Satisfying user-specified constraints while producing an answer — a prerequisite capability for agent dialogue and long-horizon task execution.
  • Research coverage: generalization mechanisms and ceiling improvement of instruction following under RL (ACL 2026, ICLR 2027 under review), multimodal instruction-following RL data synthesis (ACL-ARR under review), and the application of on-policy distillation (OPD) to this setting.

Publications

Publications
Authors: Yirong Zeng, Shen You, et al.
NeurIPS 2026 | CCF-A, First Author
Authors: Yufei Liu*, Yirong Zeng* (co-first), et al.
NeurIPS 2026 | CCF-A, Co-First Author
Authors: Yirong Zeng, Xiao Ding, Yufei Liu, Yuxian Wang, Qunyao Du, Yutai Hou, Wu Ning, Haonan Song, Duyu Tang, Dandan Tu, et al.
ICLR 2026 | CCF-A, First Author
Authors: Yirong Zeng, Yufei Liu, Xiao Ding, Yutai Hou, Yuxian Wang, Wu Ning, Haonan Song, Dandan Tu, Qixun Zhang, et al.
ACL 2026 | CCF-A, First Author
Authors: Yirong Zeng, Xiao Ding, Bibo Cai, Bing Qin, Ting Liu
IPM 2026 (IF=8.3) | SCI-Q1, First Author
Authors: Yirong Zeng, Xiao Ding, Yutai Hou, Yuxian Wang, Li Du, Juyi Dai, Qiuyang Ding, Duyu Tang, Dandan Tu, Weiwen Liu, Bing Qin, Ting Liu
EMNLP 2025 | Findings, CCF-B, First Author
Authors: Yirong Zeng, Xiao Ding, Yuxian Wang, Weiwen Liu, Wu Ning, Xu Huang, Duyu Tang, Dandan Tu, Bing Qin, Ting Liu
EMNLP 2025 | Main, CCF-B, First Author
Authors: Yirong Zeng, Xiao Ding, Bibo Cai, Ting Liu and Bing Qin
NAACL 2025 | Main, CCF-B, First Author
Authors: Yirong Zeng, Xiao Ding, Yi Zhao, Xiangyu Li, Jie Zhang, Chao Yao, Ting Liu and Bing Qin
COLING 2024 | CCF-B, First Author
Less Harness, More Signal: Efficient In-Harness RL for Autonomous Agents.
Authors: Yufei Liu*, Yirong Zeng* (co-first), et al.
AAAI 2027 | Under Review, CCF-A
Authors: Yirong Zeng, Yufei Liu, et al.
ICLR 2027 | Under Review, CCF-A
Authors: Yirong Zeng, Shen You, et al.
ICLR 2027 | Under Review, CCF-A
Constraint-Decoupled On-Policy Distillation for Multi-Constraint Instruction Following.
Authors: Yirong Zeng, Weian Yu, et al.
ICLR 2027 | Under Review, CCF-A
Not Every Tool Call Helps: Fine-Grained Credit Assignment for Efficient Tool-Integrated Reasoning.
Authors: Shen You*, Yirong Zeng* (co-first), et al.
ICLR 2027 | Under Review, CCF-A
Authors: Yirong Zeng, Sai Zhang, et al.
ACL-ARR | Under Review, CCF-A

Industry Experience

Industry Experience

AI Algorithm Engineer (Research Intern)

Huawei (Beijing) · Xiaoyi Foundation Algorithm Dept. | 2024.09 - 2026.09

As a research intern, led the pioneer exploration of system-level post-training techniques for Huawei Xiaoyi’s large models: 12 research outcomes across Agentic RL, Harness Agent training and instruction following, 4 of which shipped in Xiaoyi product releases. I was among the first batch of the Huawei Xiaoyi Business Unit Top-Talent Program interns and served as the intern lead of the department.

Software & Datasets

Software & Datasets
ClawLoop Lightweight Harness & RL Training Framework
Yirong Zeng, et al. · ⭐ 33 · GitHub
ClawForge (EnvCraft) Agentic RL Environments, Training Code & Models
Yirong Zeng, et al. · ⭐ 91 · GitHub
MIFS Multimodal Instruction-Following Dataset
Yirong Zeng, et al. · 11k+ downloads
Misinformation Detection Real-time Rumor Identification & Evidence Tracing
Yirong Zeng, Juyi Dai, Shen You · Demo System