Reproduce “Language Models can Solve Computer Tasks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Geunwoo Kim, Pierre Baldi, Stephen McAleer
A novel Recursive Criticism and Improvement (RCI) prompting scheme enables pre-trained Large Language Models (LLMs) to effectively solve general computer tasks, guided solely by natural language commands. The approach achieves state-of-the-art performance on the MiniWoB++ benchmark using fewer than two demonstrations per task and also enhances general LLM reasoning capabilities.
Key insights
- LLMs can effectively self-critique and iteratively refine their outputs to address the task, state, and agent grounding challenges inherent in computer interaction.
- The RCI prompting scheme dramatically reduces data requirements for computer task automation, achieving high performance with unprecedented sample efficiency compared to traditional methods.
- RCI is a general-purpose technique that also enhances LLMs' reasoning abilities on natural language tasks, synergizing with existing methods like Chain-of-Thought.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.