At 10:00 a.m. on May 30, Dean Dai Wangzhou of our institute invited Dr. Yu Gu from The Ohio State University to give an academic report titled “Extrapolating A Agents to Open-World Tasks with Natural Language.”
Abstract:
Enabling machines to take actions and fulfill diverse human goals constitutes, arguably, the most longstanding objective of Al research. A core challenge for developing such versatile agents in open-worlctasks lies in effective goal transfer (i.e., extrapolating to diverse unseen goals). Agents based onreinforcement learning (RL) specify goals as reward functions and are typically only optimized toachieving a fixed goal (e.g., playing Go or cleaning a table). Extrapolating to new goals would requirespecifying new reward functions and optimization against them. We advocate for explicitly modelinggoals in natural language in Al agents, This not only creates a friendly interface between humans ancagents, but more importantly, it also allowsextrapolate to new goals more easily, In this talkwe will present an intuitive high-level view o.how natural anguage modeling facilitates extrapolationand discuss several proiects under this framework, In addition, Dr, Yu Gu will also share some personamusings regarding agent research in the LLM era and Al research in general
Speaker's Profile:
Dr. Yu Gu received his Ph.D. from The Ohio State University and obtained his Master's and Bachelor'degrees from the Department of Computer Science at Nanjing University. His primary research focuseson language agents, with a long-standing commitment to exploring the critical role of language irenabling agents to generalize across diverse tasks in open-world settings. Additionally, he has broadinterests in neuro-symbolic Al and cognitively inspired Al models. His work has been published inseveral prestigious conferences, including ACL, EMNLP, NAACL, COLING, NeurlPS. and WWW, Hehas received Outstanding Paper Awards as the first author at both ACL and COLING.
