Odyssey releases CaliBench to test if world models reproduce physical randomness
damianplayer · x · 2026-08-19
Odyssey introduced CaliBench, a benchmark to evaluate if video world models reproduce the true stochasticity of the universe, such as dice roll distributions, rather than just generating visually plausible frames. The post highlights that current top video models perform poorly at calibration—accurately reproducing statistical distributions of physical outcomes like Galton board results.
More from Research
- JointDiff: one multimodal diffusion model maps fusion experiments to physics both ways — bravo_abad · 2026-08-20
- AntLing open-sources Ling-3.0 models using WSM to replace LR decay — AcanthisittaOk1699 · 2026-08-19
- VGGT-Align anchors scene geometric invariants to fix scale drift in long-sequence 3D reconstruction — zhenjun_zhao · 2026-08-19
- SplatGuide reuses one 3DGS reconstruction as three priors, hitting SOTA pose-free novel view synthesis — zhenjun_zhao · 2026-08-19
- UniQuery4R encodes a clip once and unifies 4D scene reconstruction via query-conditioned decoding — zhenjun_zhao · 2026-08-19
- GeoWeaver couples a geometric prior model with test-time adaptation to fix long-video 3D reconstruction drift — zhenjun_zhao · 2026-08-19