I am a second-year master's student in Computer Science and Technology at ShanghaiTech University, advised by Prof. Kan Ren. I am also an LLM Post-Training Research Intern at Ant Group.
Kun Feng*, Yuchen Fang*, Yiyang Tan, Shuqi Gu, Yongxiang Zhao, Yu Liu, Xingyu Lu, Lintao Ma, Kan Ren†
As long-horizon agents improve, static evaluation criteria and training priorities can fall out of sync with evolving capability gaps; sparse feedback and repeated failures further hinder learning. ARISE co-evolves rubrics and skills with capability-based adaptive sampling to align evaluation, exploration, and training with changing learning needs.随着长程智能体能力提升，固定的评价标准与训练重点可能逐渐偏离不断变化的能力缺口；稀疏反馈与反复失败进一步阻碍学习。ARISE 通过评估准则与技能协同演化，以及基于能力的自适应采样，使评价、探索与训练持续适应变化的学习需求。
Kun Feng*, Ziwei Shan*, Yuchen Fang, Yiyang Tan, Sihan Lu, Shuqi Gu, Xingyu Lu, Lintao Ma, Kan Ren†
Multimodal forecasting requires both semantic reasoning and numerical precision, yet LLMs struggle with accurate quantitative prediction while time series foundation models often lack future-oriented semantic reasoning. KairosAgent bridges this gap by fusing tool-augmented LLM reasoning into a time series forecaster and refining reasoning through forecasting rewards.多模态预测需要兼顾语义推理与数值精度，但大语言模型难以进行精确的定量预测，时间序列基础模型则往往缺乏面向未来的语义推理能力。KairosAgent 将工具增强的大语言模型推理融入时间序列预测器，并通过预测奖励优化推理，弥合两者之间的鸿沟。
Kun Feng*, Shaocheng Lan*, Yuchen Fang*, Wenchao He, Sihan Lu, Shuqi Gu, Lintao Ma, Xingyu Lu, Kan Ren†
Varying sampling densities and periodic structures challenge zero-shot forecasting: static tokenization and positional encoding entangle diverse temporal patterns, driving reliance on larger models. Kairos adapts patch sizes to local information density and positional encodings to instance-specific temporal structure, enabling generalization with substantially fewer parameters.采样密度与周期结构的差异给零样本预测带来挑战：静态分块与位置编码将多样的时间模式混杂在固定表示中，使模型依赖更大的参数量。Kairos 根据局部信息密度自适应选择分块粒度，并依据实例的时间结构调整位置编码，以显著更少的参数实现泛化。
ARISE · Agentic reinforcement learning. Developed a framework that co-evolves evaluation rubrics and guiding skills, with capability-based adaptive task sampling, to address evolving capability gaps in long-horizon agents. Improved task performance and training efficiency on SkillsBench and Terminal-Bench.ARISE · 智能体强化学习。围绕长程智能体持续变化的能力缺口，设计评估准则与指导技能协同演化、基于能力的自适应任务采样框架，提升 SkillsBench 和 Terminal-Bench 上的任务表现与训练效率。
KairosAgent · Agentic forecasting. Combined tool-augmented LLM reasoning with a time series foundation model through gated cross-modal fusion. Built the T-STAR reasoning corpus and a post-training pipeline with SFT, forecaster alignment, and GRPO using turn-level forecasting rewards.KairosAgent · 智能体预测。通过门控跨模态融合，将工具增强的大语言模型推理与时间序列基础模型结合。构建 T-STAR 推理轨迹语料，并搭建包含 SFT、预测器对齐及基于轮次级预测奖励的 GRPO 后训练流程。
Earned six national-level and six provincial-level awards in competitions including the Group Programming Ladder Tournament, Lanqiao Cup, Mathematical Contest in Modeling (MCM), and Jiangsu Provincial Higher Mathematics Competition. CCF Computer Software Proficiency Certification (CSP): 440/500 (top 0.27% in the cumulative ranking).累计获得六项国家级奖项和六项省级奖项，包括团体程序设计天梯赛、蓝桥杯、美国大学生数学建模竞赛、江苏省高等数学竞赛。 CCF计算机软件能力认证：440/500分（累计排名为前0.27%）。
//Academic Service学术服务
ICLR 2027
Reviewer审稿人
//Teaching教学经历
Machine Learning机器学习
Teaching Assistant助教
Fall 20262026 年秋季
AI FOR FACULTY (School of Entrepreneurship and Management)AI FOR FACULTY（创管学院）
Teaching Assistant助教
Fall 20252025 年秋季
//Selected Honors & Awards部分荣誉与奖项
✧
CCF Outstanding College StudentCCF优秀大学生启航计划Top 100 nationwide全国百名
✧
National Scholarship国家奖学金Top 0.2% nationwide全国前 0.2%
✧
Qingying Feifan Scholarship青颖飞帆奖学金Sole recipient in the major专业唯一获奖者
