Speaker: Shunyu Yao Princeton University
Time: 2023-04-12 18:00-2023-04-12 19:30
Venue: C19-2 (Online：https://us06web.zoom.us/j/89924841514?pwd=WHQ0elp2eWp0YXRzbEJCaUt6YTN0Zz09)

Abstract:

Large language models have revolutionized natural language processing, but they are trained to write, not to act or make choices that interface with our world. This talk discusses two basic aspects for incorporating language models with the automacy of interacting with external environments, to learn beyond static corpora and accomplish goals beyond text generation. (1) Building environments to develop language agents. We present the WebShop benchmark, and show interaction with the Internet is more scalable than existing paradigms like games and dialog systems, and enables transfer to real-world tasks, e.g. shopping on amazon.com. (2) Adapting language models for interactive learning. We present ReAct, an idea to synergize reasoning and acting for language models, which achieve state-of-the-art few-shot results on multiple reasoning and sequential decision making problems, with reduced hallucinations and increased human alignment. We also present CALM, an idea to integrate language models with reinforcement learning agents to flexibly adapt to downstream tasks, while preserving pre-trained priors and avoiding language drift.

Short Bio:

Shunyu Yao is a 4th year Phd student with Karthik Narasimhan at Princeton NLP Group, working at the intersection of language and reinforcement learning. Previously, he graduated from Yao Class at Tsinghua University, and spent time at MIT, Microsoft, IBM, and Google as interns.