Research map / Multi-agent systems and games / LLM agents
AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
"Shows that carefully aligning a web agent's observation and action space with LLM capabilities, with no in-context examples or extra modules, sets a new WebArena state of the art."
Problem
"x"
Approach
Keep the base LLM policy fixed and only change the rule-based observation and action mappings: remove non-essential actions (noop, scroll, hover, press, tab ops), convert page content to a compact Markdown-like form, replay only pivotal-node neighborhoods and the current plan's history, and add LLM-generated branch/prune planning actions.
Key results
- AgentOccam reaches 43.1% SR on WebArena (812 tasks) with GPT-4-Turbo, versus 37.2% for WebPilot (GPT-4o) and 33.3% for SteP replication, a 15.8% relative gain over the best prior/concurrent result.
- Over the plain WebArena agent (16.5% SR), the aligned spaces add 26.6 points (+161%).
- On the 129-task WebVoyager definite-answer subset, AgentOccam reaches 54.30% vs Agent-E's 51.90%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Align the observation and action spaces | WebArena-replication with GPT-4-Turbo · WebArena | success rate (%): 16.5 -> 43.1 | Helped |
| Replace WebPilot with AgentOccam | WebArena · WebArena | success rate (%): 37.2 -> 43.1 | Helped |
| Add an LLM judge to select among generated action candidates | AgentOccam · WebArena | success rate (%): 43.1 -> 45.7 | Helped |
| Add SteP task-specific strategies | AgentOccam · WebArena | success rate (%): 43.1 -> 41.1 | Hurt |
| Remove non-essential and embodiment-heavy actions (ablation) | AgentOccam with GPT-4-Turbo · WebArena development set | success rate (%): 14.2 -> 25.8 | Helped |
| Disable scrolling and provide the full page (ablation) | AgentOccam with GPT-4-Turbo · WebArena development set | success rate (%): 25.8 -> 30.0 | Helped |
| Simplify web page elements and remove redundant text (ablation) | AgentOccam with GPT-4-Turbo · WebArena development set | success rate (%): 30.0 -> 34.7 | Helped |
| Add selective replay of pivotal web elements in history (ablation) | AgentOccam with GPT-4-Turbo · WebArena development set | success rate (%): 34.7 -> 36.8 | Helped |
| Add selective replay of pivotal web elements in history (ablation) | AgentOccam · WebArena Shopping and Reddit | success rate (absolute points): 3.2 and 6.0 points | Hurt |
| Add branch/prune planning and selective replay of past plans (ablation) | AgentOccam with GPT-4-Turbo · WebArena development set | success rate (%): 36.8 -> 44.2 | Helped |
| Add branch/prune planning and selective replay of past plans (ablation) | AgentOccam with Gemini-1.5-Flash · WebArena development set | success rate (%): 32.1 -> 33.7 | Helped |
| Use AgentOccam instead of Agent-E | WebVoyager definite-answer subset · WebVoyager | success rate (%): 51.90% -> 54.30% | Helped |
Surprising findings
- Removing scrolling and expanding full pages increased observation tokens yet still helped, showing agents looped on scroll actions.
- Adding SteP's hand-written task strategies to AgentOccam lowered SR from 43.1% to 41.1%.
Limitations
- Relies on proprietary GPT-4-Turbo API; results with open models are limited to a development set with Gemini-1.5-Flash.
- Results on WebVoyager are on a small 129-task subset with human-assessed or hard-coded evaluators and single runs.
- Some task strategies (e.g. LLM judge) add extra calls, and the judge variant gives only modest gains.
Experimental setup
Datasets: WebArena WebVoyager
Benchmarks: WebArena WebVoyager
Metrics: success rate observation tokens per step steps per task
Baselines: WebArena CoT agent (replication) SteP (replication) WebPilot Agent Workflow Memory Agent-E
Builds on: WebArena SteP Agent Workflow Memory
Closest papers by meaning
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.90)
- How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.90)
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.90)
- A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.89)
- WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.89)
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.