'Execution Before Memory' Paper: Video World Models Fail at Execution, Not Memory

The MIST evaluation framework separates video-world-model memory testing into scene setup, action execution and final state reveal, and finds the main failure mode is execution rather than memory persistence. The authors' M3-Agent recursively revises instructions on a frozen base model, lifting scores 26.1% on WRBench and 52.7% on MBench-T.