时 间:2026年11月15 日(周日)14:00 - 16:00
地 点:普陀校区理科大楼A614室
报告人:吕晨阳 阿里巴巴集团高级研究员
主持人:李文熙 华东师范大学统计学院
摘 要:
How can we tell whether an AI agent is improving its ability to improve? This talk presents DeltaHarness, an experimental framework for studying recursive self-improvement in tool-using agents. DeltaHarness separates execution programs (H), decisions during execution (Q), and experiment-selection policies (I), while keeping the proposing language model’s weights fixed. Using real media-processing workflows, we examine physical plan optimization, feedback-driven adaptation, and changes to the improvement process. The current evidence shows that fixed reuse plans can reduce execution cost and that LLM-generated policies can outperform strong fixed plans. However, additional gains from iterative feedback and recursive improvement remain unestablished in the reported studies. I will discuss what these results reveal about feedback timing, policy validity, resource variation, and the distinction between execution savings and the cost of obtaining them. The talk concludes with a registered transfer protocol for evaluating whether improvements persist across new tasks and resource conditions.
报告人简介:
Chenyang Lyu 现任阿里巴巴集团 ATH-MaaS AI 高级研究员兼技术负责人,领导语音与全模态大语言模型应用研究团队。他的研究聚焦于多语言与多模态大语言模型、语音智能以及可靠的 AI 智能体,已在 NeurIPS、ICML、ACL、EMNLP、CVPR 和 ACM Multimedia 等国际顶级会议发表论文 40 余篇。他的代表性工作包括 Marco 模型系列,以及面向长视频编辑的智能体框架 Crayotter。 他的研究曾获 ACL 2026 SAC Highlights 奖及 ACM Multimedia 2024 最佳论文奖提名。他曾获爱尔兰 2023 年 AI Awards“年度AI青年”称号,并获选德国 DAAD AInet Fellow。他带领的团队还在 IWSLT 2025 中获得两项冠军和两项亚军。 加入阿里巴巴之前,他曾任穆罕默德·本·扎耶德人工智能大学研究员,并于 2023 年获得都柏林城市大学博士学位。他目前的研究兴趣包括全模态大语言模型、智能体的可靠性与记忆,以及递归自我改进。