Toward Memory-Aided World Models: Benchmarking via Spatial Consistency

The project addresses a critical bottleneck in current generative AI: the lack of physical and spatial 'common sense' in video and world models (the 'Sora' problem). With 0 stars and 4 forks only 9 days after release, this is an early-stage academic artifact tied to a specific paper. Its defensibility is currently low as it functions primarily as a set of evaluation scripts and metrics rather than a production-grade tool. However, the focus on 'spatial consistency via memory' is a high-interest niche for frontier labs (OpenAI, Google DeepMind, Meta FAIR) who are currently struggling with model 'hallucinations' in physical space. While frontier labs will likely develop proprietary versions of these benchmarks, the methodology here could influence the standard way spatial drift is measured. The lack of traction suggests it has not yet become a community standard. The main risk is displacement by more comprehensive benchmarks from larger labs (e.g., V-JEPA's evaluation suites or future iterations of Ego4D benchmarks).