New Interactive Reward Agent Achieves 86.9% Accuracy in GUI Task Evaluation

Researchers have developed an Interactive Reward Agent (IRA) designed to improve the evaluation of GUI agents. This agent utilizes a propose-then-verify framework to gather and confirm evidence from the environment's state, combining visible interface data with system configurations and file data. The IRA achieved 86.9% accuracy on the newly introduced GUI-RewardBench, a benchmark comprising 321 GUI task trajectories across 10 Ubuntu desktop application categories. Furthermore, when applied to reinforcement learning for GUI agents, the IRA facilitated a 34.0% success rate in OSWorld, demonstrating its efficacy in providing training reward signals. AI

IMPACT Introduces a novel method for evaluating GUI agents, potentially improving their training and performance.

RANK_REASON The item is a research paper detailing a new method and benchmark for evaluating GUI agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• CORE Recommender
• DagsHub
• Gotit.pub
• GUI-RewardBench
• Hugging Face
• Influence Flower
• Interactive Reward Agent
• OSWorld
• ScienceCast
• Ubuntu

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →