NEWS · RESEARCH · #152
OdoBot: Token-efficient web-agent architecture via application behavior modeling
An arXiv preprint introduces OdoBot, a web-agent architecture that builds a behavioral model of a web application from successful task-execution demonstrations to reduce token usage. In experiments on 45 tasks in the Canvas LMS, OdoBot used 44% and 80% fewer tokens than Agent-E and WebVoyager respectively, and it outperformed WebVoyager on task success rate.
KEY POINTS
- An arXiv preprint introduces OdoBot, a web-agent architecture that builds a behavioral model of a web application from successful task-execution demonstrations to reduce token usage.
- In experiments on 45 tasks in the Canvas LMS, OdoBot used 44% and 80% fewer tokens than Agent-E and WebVoyager respectively, and it outperformed WebVoyager on task success rate.
- Lowering token consumption while maintaining or improving task success can materially reduce deployment costs and improve scalability of web agents.
WHY IT MATTERS
Lowering token consumption while maintaining or improving task success can materially reduce deployment costs and improve scalability of web agents.