4dcodebench (4dcodebench.com)

4DCodeBench has emerged as a groundbreaking framework for evaluating AI models in the realm of dynamic scene reconstruction. The competitive leaderboard showcases GPT-6 Astra [Max] and Claude Opus 5.5 [High] at the top, underscoring the ongoing dominance of proprietary models over open-weight counterparts in the quality of visual reconstruction. The benchmark assesses model performance across five metric families—appearance, geometry, and motion—using a novel method where agents generate code to reconstruct events from input videos without prior object or camera information. This approach allows researchers to differentiate the strengths and weaknesses of various models, with notable findings indicating that while static scenes are relatively well reconstructed, the reconstruction of motion remains a significant challenge. The implications of these results are substantial for the AI/ML community, as they highlight both the advances and limitations of current models in understanding complex dynamics. As demonstrated, methods such as analytic motion and custom simulations are deployed variably across models, influencing their reconstruction abilities. Furthermore, the strong correlation between model rankings and human preferences signals the reliability of automated assessments in this domain. The findings from 4DCodeBench pave the way for enhanced model development, emphasizing the need for improved methodologies to address the complexities inherent in dynamic scene reconstruction.