Jiahang Lin

Jiahang Lin林佳航

M.S. Student at FudanNLP Lab, Fudan University复旦大学 FudanNLP 实验室 硕士研究生

I am a master’s student at FudanNLP Lab, Fudan University, co-advised by Associate Prof. Tao Gui (桂 韬) and Prof. Xuanjing Huang (黄萱菁). My research focuses on self-evolving agents and reinforcement learning. I have end-to-end experience across the agent post-training stack — SFT (DeepSpeed, LLaMA-Factory), RL (verl, trl), and Harness Engineering. I am currently a Research Intern at NEX-AGI (Shanghai Qiji Zhifeng Co., Ltd.), and previously interned at the Computing Research Department of Huawei. I received my B.S. from Fudan University in 2025.我是复旦大学 FudanNLP 实验室的硕士研究生，师从桂韬副教授与黄萱菁教授。我的研究方向为自演化智能体（self-evolving agents）与强化学习。我在智能体后训练全链路上拥有端到端经验——SFT（DeepSpeed、LLaMA-Factory）、RL（verl、trl）以及 Harness 工程。我目前在 NEX-AGI（上海奇绩智峰）担任研究实习生，此前曾在华为计算研究部门实习。我于 2025 年在复旦大学获得学士学位。

You can find my publications on Google Scholar. Feel free to reach out to me at 1187524561@qq.com.你可以在 Google Scholar 上查看我的论文。欢迎通过邮箱 1187524561@qq.com 与我联系。

Total GitHub Stars 1st Prize · Tencent Competition Domestic Finals

🔥 News最新动态

  • 2026.05: 🎉🎉 AHE (Agentic Harness Engineering) is released on arXiv!🎉🎉 AHE（Agentic Harness Engineering）已在 arXiv 发布！
  • 2026.04: 🎉🎉 MM-Doc-R1 is accepted by ACL 2026 Findings!🎉🎉 MM-Doc-R1 被 ACL 2026 Findings 录用！
  • 2026.04: 🎉🎉 OctoBench is accepted by ACL 2026 Main!🎉🎉 OctoBench 被 ACL 2026 主会 录用！
  • 2025.12: 🎉🎉 Nex-N1, our agentic model trained via a unified ecosystem for large-scale environment construction, is released on arXiv!🎉🎉 Nex-N1——我们通过统一生态进行大规模环境构建所训练的智能体模型，已在 arXiv 发布！

💻 Internships实习经历

  • NEX-AGI2025.3 - Present · Research Intern, NEX-AGI (Shanghai Qiji Zhifeng Co., Ltd.), Shanghai2025.3 - 至今 · 研究实习生，NEX-AGI（上海奇绩智峰），上海
    Self-evolving agents & reinforcement learning for coding and search agents面向代码与搜索智能体的自演化智能体与强化学习
  • Huawei2024.7 - 2024.9 · Research Intern, Huawei Computing Product Line, Computing Research Department2024.7 - 2024.9 · 研究实习生，华为 计算产品线 计算研究部
    Fine-tuning long-context LLMs on Ascend 910B在昇腾 910B 上微调长上下文大模型
  • FudanNLP
    2024.4 - 2024.7 · Research Intern, FudanNLP Lab, Fudan University2024.4 - 2024.7 · 研究实习生，FudanNLP 实验室，复旦大学
    LLM safety: training red-team models for attack and defense大模型安全：训练用于攻防的红队模型

📝 Publications论文发表

Arxiv 2026
sym

Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses

Jiahang Lin*, Shichun Liu*, Chengjun Pan*, Lizhi Lin, Shihan Dou, Xuanjing Huang, Hang Yan, Zhenhua Han†, Tao Gui†

  • AHE is an observability stack for the automatic optimization of coding-agent harnesses, with three pillars: component observability (NexAU), experience observability (Agent Debugger), and decision observability (evidence-driven Evolve Agent).AHE 是一套用于自动优化编码智能体 harness 的可观测性体系，包含三大支柱：组件可观测性（NexAU）、经验可观测性（Agent Debugger）以及决策可观测性（基于证据的 Evolve Agent）。
  • Without changing the model, AHE pushes Terminal-bench 2 from 69.7% to 77.0% across iterations, with strong cross-task and cross-model generalization.在不改动模型的前提下，AHE 通过多轮迭代将 Terminal-bench 2 从 69.7% 提升至 77.0%，并展现出强的跨任务、跨模型泛化能力。
  • | | |
Arxiv 2026
sym

EVPO: Explained Variance Policy Optimization for Adaptive Critic Utilization in LLM Post-Training

Chengjun Pan*, Shichun Liu*, Jiahang Lin*, Dingwei Zhu, Jiazheng Zhang, Shihan Dou, Songyang Gao, Zhenhua Han, Binghai Wang, Rui Zheng, Xuanjing Huang†, Tao Gui†, Yansong Feng†

  • We cast baseline selection in LLM post-training as a Kalman filtering problem, unifying PPO and GRPO as two extremes of the Kalman gain, and prove that the sign of explained variance (EV) is the exact boundary separating the variance-reducing from the variance-inflating critic regime.我们将大模型后训练中的 baseline 选择建模为卡尔曼滤波问题，把 PPO 与 GRPO 统一为卡尔曼增益的两个极端，并证明 explained variance（EV）的符号正是区分「降方差」与「增方差」critic 区间的精确边界。
  • EVPO adaptively switches between critic-based and batch-mean advantage estimation per step based on EV sign, achieving the best results across Sokoban, FrozenLake, WebShop, and MATH.EVPO 依据每一步的 EV 符号，在基于 critic 与基于 batch 均值的优势估计之间自适应切换，在 Sokoban、FrozenLake、WebShop 与 MATH 上均取得最佳结果。
  • |
Arxiv 2026
sym

MM-Doc-R1: Training Agents for Long Document Visual Question Answering through Multi-turn Reinforcement Learning

Jiahang Lin*, Kai Hu*, Binghai Wang, Yuhao Zhou, Zhiheng Xi, Honglin Guo, Shichun Liu, Junzhe Wang, Shihan Dou, Enyu Zhou, Hang Yan, Zhenhua Han, Tao Gui†, Qi Zhang†, Xuanjing Huang†

  • Conventional RAG systems struggle with complex multi-hop queries over long documents due to their single-pass retrieval. MM-Doc-R1 trains agents for long-document visual question answering via multi-turn reinforcement learning.传统 RAG 系统受限于单次检索，难以应对长文档上的复杂多跳查询。MM-Doc-R1 通过多轮强化学习训练面向长文档视觉问答的智能体。
  • The agent learns to interleave retrieval and reasoning across many turns, achieving substantial gains on long-document VQA benchmarks compared to single-pass and prompt-only baselines.该智能体学会在多轮之间交替进行检索与推理，相比单次检索与仅提示（prompt-only）基线，在长文档 VQA 基准上取得显著提升。
Arxiv 2026
sym

OctoBench: Benchmarking Scaffold-Aware Instruction Following in Repository-Grounded Agentic Coding

Deming Ding*, Shichun Liu*, Enhui Yang*, Jiahang Lin*, Ziying Chen, Shihan Dou, Honglin Guo, Weiyu Cheng, Pengyu Zhao, Chengjun Xiao, Qunhong Zeng, Qi Zhang, Xuanjing Huang, Qidi Xu, Tao Gui

  • Modern coding scaffolds turn LLMs into capable software agents, but their ability to follow scaffold-specified instructions remains under-examined. OctoBench benchmarks scaffold-aware instruction following in repository-grounded agentic coding.现代编码脚手架（scaffold）能把大模型变成强大的软件智能体，但它们遵循脚手架指定指令的能力仍缺乏系统评估。OctoBench 在仓库级智能体编码场景下，评测面向脚手架的指令遵循能力。
  • OctoBench includes 34 environments and 217 tasks under three scaffold types, with 7,098 objective checklist items. We release an automated observation-and-scoring toolkit for full trajectories and fine-grained checks.OctoBench 包含三类脚手架下的 34 个环境、217 个任务，以及 7,098 条客观检查项。我们开源了一套可对完整轨迹进行自动观测与打分、支持细粒度检查的工具包。
  • | | |
Arxiv 2025
sym

Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

Nex-AGI Team: Yuxuan Cai, Lu Chen, …, Jiahang Lin, …, Xuanjing Huang, Xipeng Qiu

  • We introduce a comprehensive method designed to systematically scale the diversity and complexity of interactive environments through three orthogonal dimensions: Complexity (NexAU), Diversity (NexA4A), and Fidelity (NexGAP).我们提出一套系统性方法，从三个正交维度规模化提升交互环境的多样性与复杂度：复杂度（NexAU）、多样性（NexA4A）与保真度（NexGAP）。
  • Nex-N1 consistently outperforms SOTA open-source models and achieves competitive performance against frontier proprietary models on complex agentic tasks (SWE-bench, tau2).在复杂智能体任务（SWE-bench、tau2）上，Nex-N1 持续超越 SOTA 开源模型，并在与前沿闭源模型的对比中展现出有竞争力的表现。
  • |

🎖 Honors and Awards荣誉奖项

  • 2024, 1st Prize, 2024 Tencent AI Arena Global Open Competition · Agent Game Algorithm Track · Mainland China Regional Final. Team: 五角场三分王, Fudan University.2024，一等奖，2024 腾讯 AI Arena 全球公开赛 · 智能体博弈算法赛道 · 中国大陆赛区总决赛。队伍：五角场三分王，复旦大学。
  • 2022-2023, Third Prize, Fudan University Outstanding Student Scholarship.2022-2023 学年，复旦大学优秀学生奖学金三等奖。
  • 2021-2022, “Star of Yangfan” of Fujian Province; Youth Talent Recruitment Volunteer of Fujian Province.2021-2022 学年，福建省扬帆之星；福建省青年引才志愿者。

📖 Education教育经历

  • 2025.09 - Present, M.S., Fudan University (FudanNLP Lab).2025.09 - 至今，硕士，复旦大学（FudanNLP 实验室）。
  • 2021.09 - 2025.06, B.S., Fudan University.2021.09 - 2025.06，学士，复旦大学。