Please note: This PhD seminar will take place online.
Dongfu Jiang, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Wenhu Chen
Recent progress in reinforcement learning with verifiable rewards has substantially improved the reasoning capabilities of large language models, but most existing training paradigms still assume a closed, single-turn setting. In many realistic tasks, however, strong reasoning requires interaction with external tools such as code interpreters, search engines, SQL executors, web browsers, and vision utilities.
This talk presents VerlTool, a unified and extensible framework for agentic reinforcement learning with tool use. VerlTool formulates tool-augmented learning as multi-turn trajectories with external observations, decouples reinforcement learning workflows from tool execution through a standardized tool-server interface, and supports asynchronous rollouts to reduce bottlenecks during training. I will discuss the motivation behind agentic RL with tools, the systems challenges that arise when scaling multi-turn tool interaction, and how VerlTool addresses fragmentation, extensibility, and efficiency across diverse domains including math reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering. The talk will conclude with broader lessons on building infrastructure for future tool-using LLM agents.