arXiv cs.AIMonday · June 1, 2026FREE

Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration

Researchers propose SCALE (Self-Cognitive-Aware Learning and Exploration), a framework for building web agents that improve themselves through exploration. SCALE uses three adversarial roles: Selector, Predictor, and Judger. The Selector identifies actions, the Predictor anticipates outcomes, and the Judger evaluates success, enabling the agent to detect its own limitations and learn from mistakes. To avoid local exploration traps, SCALE-Hop employs graph-based global planning. The team also released SCALE-20k, a dataset of structured demonstrations from 19 real-world websites covering diverse tasks. Experiments show that SCALE significantly boosts performance and generalization of multiple Multimodal Large Language Models (MLLMs) in various web environments. The framework offers a scalable solution for building adaptive web agents without relying on handcrafted pipelines or expensive expert trajectories.