SteP uses stacked LLM policies to improve web task performance

Researchers propose SteP (Stacked LLM Policies for Web Actions), an approach that dynamically composes policies to solve diverse web tasks by defining a Markov Decision Process where the state is a stack of policies. Unlike traditional methods restricted to static hierarchies, this method enables dynamic control that adapts to task complexity.

• On WebArena, SteP improves performance from 14.9% to 33.5% over SOTA methods using GPT-4 policies.
• On MiniWob++, SteP is competitive with prior works while requiring significantly less data.
• The approach addresses challenges like combinatorially large open-world tasks and behavior leaks by decomposing control into distinct, dynamically called policies.

The authors consider this important because it allows LLMs to handle complex web interfaces without the complexity of specifying large prompts for all possible behaviors.