BVB benchmark grades video understanding by Blender reconstruction; open-source Inkling 2 Small hits 71.6

SonglinYang4 · x · 2026-11-13

• Researchers from University of Rochester, CMU, Sony and others released BVB, a benchmark that tests agentic video understanding through reconstruction: the agent watches a real indoor video and writes Python code for Blender to rebuild it as an animated 3D scene with a camera following the original trajectory.
• All agents see only video frames, run in the same Mini-BVB sandbox, and face identical per-scene cost limits; reconstructions are scored against the source via Dual VQA and Latent Similarity.
• Claude Opus 5.2 high (76.4) and GPT-6 Astra high (76.3) lead the board; thinkymachines' newly open-sourced (Apache 2.0) Inkling 2 Small (max) debuted at 71.6, a big jump from Inkling Small's 59.0, beating Gemini 3.1 Pro and DeepSeek v5.1.

Original post →

More from Models

• Microsoft diagnostic AI solves 83% of NEJM cases vs 36% for physicians — realmeetjames · 2026-11-13
• Opus 5.5 generates a full Blender 3D model in seconds, wows game developer — Scobleizer · 2026-11-13
• Claude's JavaScript art demo draws praise from threepointone: 'just remarkable' — threepointone · 2026-11-13
• Leaked UI shows unreleased Google models 'Hydrogen' and 'Argon' alongside Gemini tiers — Gauri_the_great · 2026-11-13
• DeepMind researcher: AI Overviews is the world's most-used AI model — cgarciae88 · 2026-11-13
• Tencent Hy4 Tops CWE-Bench-AA, Ranks #3 on Artificial Analysis Cyber Index — ArtificialAnlys · 2026-11-13

Models channel →