Reproduce “Web-CogReasoner: Towards Multimodal Knowledge-Induced Cognitive Reasoning for Web Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuhan Guo, Cong Guo, Aiwen Sun, Hongliang He, Xinyu Yang, Yue Lu, Yingji Zhang, Xuntao Guo and 7 more
Researchers from multiple institutions developed Web-CogReasoner, a multimodal web agent that systematically acquires factual, conceptual, and procedural knowledge inspired by Bloom's Taxonomy to enable cognitive reasoning. The agent achieved an 84.4% overall accuracy on the Web-CogBench and a 30.2% success rate on live web tasks in WebVoyager, surpassing various baselines.
Key insights
- Systematic, hierarchical knowledge acquisition (Factual, Conceptual, Procedural), guided by an educational framework, significantly enhances web agents' cognitive abilities and generalization across diverse tasks.
- A Knowledge-driven Chain-of-Thought (KCoT) reasoning mechanism is crucial for dynamically activating and applying learned knowledge, leading to more informed decision-making in complex web interactions.
- Curriculum learning, which fine-tunes models incrementally from simple to complex tasks, effectively promotes better understanding and prevents premature task termination in web agents, contributing to robust performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.