Dual Exam Benchmark Decouples Multimodal Agent and World Model Evaluation
October 23, 2026
Dual Exam separates agent action selection from environment state evolution to evaluate multimodal agents and world models independently. Testing 10 agent configurations and 15 world model configurations across physical and digital domains, results show GPT 6 Astra leading the agent index with a score of 31.3.
HOW THIS AFFECTS YOU
●
researcherYou can use this to isolate whether failures in autonomous systems stem from decision-making or environmental simulation.