4DCodeBench: strong static reconstruction doesn't transfer to dynamic scenes
jwt0625 · x · 2026-10-13
4DCodeBench (arXiv 2610.03715) benchmarks 4D inverse graphics via code generation: agents must reconstruct dynamic scenes from video as executable graphics programs, implementing abstractions like physical simulation.
- Covers real-world videos and synthetic scenes spanning deformation, fluid flow, and fracture.
- Benchmarks 18 frontier models on 200 scenes, reporting human/VLM Elo and VQA accuracy.
Key finding: strong static reconstruction does not yet translate into reliable reconstruction of complex dynamics, making it a testbed for tracking code-based world-dynamics understanding.
Related event: 4DCodeBench: Benchmarking AI Agents on 4D Scene Reconstruction via Code(2 posts)→
More from Research
- François Fleuret on why matrix multiplication keeps GPUs perfectly busy — francoisfleuret · 2026-10-13
- Specula has found 500+ bugs in distributed systems; Apache Ozone adopts it — tianyin_xu · 2026-10-13
- Researchers debate whether an n(log n)^a integer-multiplication result was misread — rrwilliams · 2026-10-13
- Reinventing the wheel: hand-rolled upscaling for tiny retro-console pixel sprites — ssh4net · 2026-10-13
- Preferred Networks' T3lescope reconstructs high-fidelity 3D scenes from multi-view images without per-scene optimization — pfnet · 2026-10-13
- ARCS benchmark: ambiguous text-to-SQL stumps models, best at just 51% accuracy — megagonlabs · 2026-10-13