WebChallenger: A Reliable and Efficient Generalist Web Agent

WebChallenger is a new web agent framework designed to enhance autonomous web navigation for LLMs by addressing cognitive gaps in existing architectures. It utilizes a structured page representation called PageMem, which organizes web content hierarchically, and incorporates three mechanisms that mimic human cognitive advantages: selective attention, persistent memory, and procedural fluency. The framework, which operates with off-the-shelf models without fine-tuning, achieves competitive benchmark scores (56.3% on WebArena, 48.7% on VisualWebArena, 51.0% on Online-Mind2Web, and 70.9% on WorkArena), making it a cost-effective alternative to proprietary systems for practitioners developing generalist web agents.