🙋

9/ Stay tuned for more updates!
🔗 Paper: arxiv.org/abs/2411.150...
🌐 Blog: http://scribehow.com/library/scribe-agent
💻 Code: github.com/colonylabs/S...
👥 Team: @junhongshen1.bsky.social Atishay Jain, Zedian Xiao, Ishan Amlekar, Mouad Hadji, Aaron Podolny @atalwalkar.bsky.social

8/ What's next? The possibilities are vast—from integrating advanced reasoning and planning modules to exploring multi-modal systems. ScribeAgent highlights the potential of production-scale training data, paving the way for future web agents that are both powerful and cost-effective.

7/ Beyond performance, ScribeAgent models also provide efficiency gains relative to most proprietary baselines, which are typically larger in size and slower at inference time. This makes ScribeAgent an attractive option in terms of accuracy, latency, and cost.

6/ Our results? ScribeAgent outperforms GPT-4o on our internal dataset and achieves state-of-the-art direct generation performance on the public benchmark Mind2Web. Our multi-agent system integrating GPT-4o also improves the best task success rate for text-only agents by 14.1% on WebArena.

5/ Combining next-step prediction with effective HTML preprocessing, we fine-tune two versions of ScribeAgent. The cost-efficient 𝗦𝗰𝗿𝗶𝗯𝗲𝗔𝗴𝗲𝗻𝘁-𝗦𝗺𝗮𝗹𝗹 is based on 7B Qwen2, while the better-performing 𝗦𝗰𝗿𝗶𝗯𝗲𝗔𝗴𝗲𝗻𝘁-𝗟𝗮𝗿𝗴𝗲 is based on 32B Qwen2.5.

4/ Data is the key! We leverage Scribe scribehow.com/, an AI documentation software that streamlines the creation of step-by-step guides for web tasks, to collect large-scale action data executed by real users on over 250 web domains. See scribehow.com/shared for example workflows.

3/ Most existing web agents rely heavily on prompting general-purpose proprietary models like GPT-4. However, LLMs like GPT-4 are not specifically trained to parse languages like HTML, limiting the agent's ability to plan and reason. In contrast, ScribeAgent adapts the LLM itself for web navigation.

2/ Web agents navigate through websites to solve real-world tasks. After the user defines a high-level objective, the agent outputs step-by-step actions based on the objective, observation, and interaction history. For text-based agents, the observation typically includes the website's URL and HTML.