DUDE Framework Teaches Web Agents to Resist Deceptive Interfaces
A team of researchers has introduced DUDE (Deceptive UI Detector & Evaluator), a two-phase framework designed to enhance the resilience of vision-language model-based web agents against misleading interface components. This framework integrates hybrid-reward learning, asymmetric penalties, and experience summarization to extract failure patterns into usable insights. Additionally, a novel benchmark named RUC (Real UI Clickboxes) has been established, featuring 1,407 scenarios spanning four domains and various deception categories. Testing indicates that DUDE decreases vulnerability to deception by 53.8% while preserving overall task performance.
Key facts
- DUDE is a two-stage framework for deception-aware web agent defense.
- It uses hybrid-reward learning with asymmetric penalties.
- Experience summarization distills failure patterns into transferable guidance.
- RUC benchmark includes 1,407 scenarios across four domains and deception categories.
- DUDE reduces deception susceptibility by 53.8%.
- Task performance is maintained with DUDE.
- The research is from the field of Artificial Intelligence.
- The paper is available on arXiv.
Entities
Institutions
- arXiv