# AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents

**Authors:** Ke Yang, Yao Liu, Sapana Chaudhary, Rasool Fakoor, Pratik Chaudhari, George Karypis, Huzefa Rangwala  
**arXiv:** [2410.13825](https://arxiv.org/abs/2410.13825)  
**Date:** 2024-10-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.13825v1/

## Key points

- Autonomous agents utilizing large language models (LLMs) boost efficiency of human interactions significantly.
- Automation of web-based tasks, like booking hotels within specified budgets, is increasingly desirable in today's digital landscape.
- Web agent serves as a proof-of-concept for various agent grounding scenarios, driving advancements in future applications.
- Existing research often involves manual crafting of strategies for web agents, which may not generalize well across real-world scenarios.
- Misalignment between a web agent's observation/action representation and the pre-training data of the LLM can be a challenge.
- Recent study focused on refining observation and action space to align with underlying LLM capabilities, resulting in improved performance.
- AgentOccam outperforms previous methods on WebArena benchmark platform without relying on in-context examples or sophisticated search strategies.
- Tuning observation and action spaces is critical for optimizing performance in LLM-based agents.

## Summary

In the realm of autonomous agents utilizing large language models (LLMs) for personalized and standardized tasks, the efficiency of human interactions is significantly boosted. The automation of web-based tasks, such as booking hotels within specified budgets, has become increasingly desirable in today's digital landscape. Not only does this web agent cater to practical needs, but it also serves as a pivotal proof-of-concept example for various agent grounding scenarios, showcasing its potential for driving advancements in numerous future applications. While previous research often involves the manual crafting of strategies for web agents, such as prompting templates, multi-agent systems, and search methods, these approaches may not always generalize well across real-world scenarios. A notable gap in existing studies lies in the misalignment between a web agent's observation/action representation and the pre-training data of the LLM upon which it is based. This discrepancy becomes particularly evident when LLMs are primarily trained for language completion tasks rather than those involving embodied navigation actions and symbolic web elements. To address this challenge, a recent study has focused on enhancing an LLM-based web agent by refining its observation and action space to better align with the capabilities of the underlying LLM. This approach has proven highly effective, enabling the base agent to outperform previous methods across a wide range of web tasks. In particular, on WebArena—a benchmark platform featuring general-purpose web interaction tasks—the newly developed AgentOccam surpasses both the previous state-of-the-art solutions and concurrent works by significant margins. Remarkably, AgentOccam achieves these impressive results without relying on in-context examples, introducing new agent roles, soliciting online feedback or employing sophisticated search strategies. Instead its success underscores the remarkable zero-shot performance capabilities of LLMs when applied to web-related tasks. This study highlights the critical importance of carefully tuning observation and action spaces to optimize performance in LLM-based agents—a key consideration for maximizing their effectiveness in practical applications moving forward.