The paper introduces MaLiang-Harness, a framework that organizes MLLM-driven visual generation into a persistent loop of construction, inspection, and revision. It addresses the Program-to-Visual (P2V) gap, where a program executes correctly but violates visual requirements. The framework uses three mechanisms: Persistent Executable Generation (PEG) state to preserve programs and task context, Traceable Generation Process (TGP) to link edits to rendered evidence, and Revision-aware Editing and (REV) for restoration and checks.
The authors evaluated 11 closed-source MLLMs on a new image (MaLiang-IBench) and 4 on a video benchmark (MaLiang-VBench). GPT-6-Astra achieved 100% generation success on both benchmarks, with 96.0% of image tasks and 76.9% of video tasks meeting all quality thresholds. The comparison revealed that general capability scores do not predict visual generation performance, as similarly scored models differed substantially in meeting visual requirements. This framework provides a systematic way to study how MLLMs translate code into visual outcomes and highlights the limitations of general benchmarks as predictors.