Dual Exam: Controlled Evaluation of Multimodal Agents and World Models

Researchers propose the Dual Exam benchmark, which evaluates multimodal agents and world models separately in a controlled manner; a 150-case coupled paired pilot achieved no task success. The agent track compares 10 configurations on 1,088 items under a fixed environment response mechanism, covering 15 reported categories; the world model track contains 400 cases across 8 scenarios, with 15 configurations covering all and 1 covering 6 scenarios, spanning physical, digital, social, and scientific domains. GPT 6 Astra leads the case-weighted agent index at 31.3, and GPT-Image-2 leads the full-coverage world model overall leaderboard at 21.7; the two sets of scores are not directly comparable due to different tasks and metrics. Results show uneven task performance and failures to retain task-relevant state, and their predictive value for coupled success or sustained mutual learning remains to be established.