World Embedding Benchmark Tests Physical Fidelity of Video Representations Across 8,000 Cases
World-Representation-Lab · hf · 2026-10-05
World-Representation-Lab introduced the World Embedding Benchmark: 8,000 controlled simulation cases from 80 families spanning fluid mechanics, solid mechanics, dynamics, and optics & electromagnetism, each pairing rendered video with simulation-derived physical annotations for text-video retrieval, physical-property regression, and multiple-choice pair classification.
Key findings:
- Pre-trained omnimodal embedding models show weak retrieval and near-chance within-family classification, yet lightweight probes recover useful physical information from frozen embeddings
- Continual contrastive training with physics-specific pairs improves retrieval/classification but degrades property regression — a trade-off between alignment and quantitative recoverability
- Retrieval-augmented generation with MiniMax-H3 using retrieved reference videos improves physical fidelity, with stronger retrievers yielding larger gains
More from Research
- Next step for FlowBank: let the workflow bank evolve through deployment — furongh · 2026-10-05
- FlowBank: NeurIPS paper reuses complementary agent workflows for 73.40 avg vs 70.40 baseline — furongh · 2026-10-05
- 7 of 9 Frontier Models Covertly Leak Credentials to Evade Oversight in Multi-Agent Systems — illinois · 2026-10-05
- EditHero: First Long-Horizon Part-Level 3D Editing Benchmark Shows LLM Agents Win — Ruihan Yu · 2026-10-05
- OctLLM Uses Sparse Octree Tokens to Hit 3D SOTA While Preserving Language Ability — Ran Dan · 2026-10-05
- Skill2Real Achieves 78.75% Success on Real-Robot Manipulation via Agentic Skill Learning — Xincheng He · 2026-10-05