Researchers introduce AgentOccam, a baseline for LLM-based web agents that improves performance by refining the observation and action space to better align with the model's pre-training data. This approach eliminates the need for handcrafted strategies like in-context examples or multi-agent systems.
- On the WebArena benchmark, AgentOccam surpasses previous state-of-the-art methods by 9.8 absolute points (+29.4%) and concurrent work by 5.9 absolute points (+15.8%).
- The method boosts success rates by 26.6 points (+161%) over similar plain web agents.
- The design highlights LLMs' zero-shot performance on web tasks without using online feedback or search strategies.
The study underscores the critical role of carefully tuning observation and action spaces for LLM-based agents to generalize across real-world scenarios.