Paper reading: Language Models can Solve Computer Tasks

发表于 更新于 分类于 本文字数： 72 阅读时长 ≈ 1 分钟

UCI/CMU, NIPS 2023

Motivation

Previous agents need a large amount of expert demonstration and task specific reward functions to be able to solve unseen tasks.

An LLM can execute computer tasks guided by NL prompts.

Focus on the MiniWoB++ environment, which is an RL-like little game environment but on web.

Strength

  • novel(maybe? not sure) Recursively-Criticize and Improve prompt format

Challenges

  • not that related so I didn’t read thoroughly.