RaDA: Retrieval-augmented Web Agent Planning with LLMs

Agents powered by large language models (LLMs) inherit important limitations, such as the restricted context length, dependency on human-engineered exemplars (e.g., for task decomposition), and insufficient generalization. To address these challenges, we propose RaDA, a novel planning method for Web agents that does not require manual exemplars, efficiently leverages the LLMs’ context, and enhances generalization. RaDA disentangles planning into two stages: for a new given task, during Retrieval-augmented Task Decomposition (RaD), it decomposes tasks into high-level sub-tasks; next, during Retrieval-augmented Action Generation (RaA), it traverses the trajectory obtained with RaD to iteratively synthesize actions based on dynamically retrieved exemplars. We compare RaDA with strong baselines covering a broad space of design choices, using both GPT-3.5 and GPT-4 as backbones; and we find consistent improvements over previous SOTA in two challenging benchmarks, Com-pWoB and Mind2Web, covering settings with different complexities. We show the contributions of RaDA via ablation studies and qualitative analysis; and we discuss the structural benefits of our more compositional design.