Principia: video models know less high-school physics than you'd think

mariyaivasileva · x · 2026-09-25

• The Principia benchmark (accepted at NeurIPS 2026 with three 5s, but no Spotlight/Oral) tests how well video generation models grasp physics.
• Core question: "What do video models know about high school physics? Less than you'd think." It covers 8 physical laws across 500+ real, carefully calibrated scenes, using relative evaluation with objective metrics.
• Building the dataset required enormous patience: calibrated experiments, verifying physical properties, even machining parts for comparability.
• The author draws a parallel to CVPR 2024's overlooked paper "Shadows Don't Lie and Lines Can't Bend".

Original post →

More from Models

• Micius-Flash: 27B Multimodal Model for Hardware and Semiconductor Agents — zhiman-ai · 2026-11-11
• Claude IP risk-check site goes viral as users audit their IPs ahead of new rules — vista8 · 2026-11-11
• Baidu says RSI auto-optimized ERNIE inference, 6.24x decode speedup on Kunxin P800 — xiaohu · 2026-11-11
• Baidu's ECHO: multi-token prediction reusing the last layer as a drafter — xiaohu · 2026-11-11
• ERNIE 5.5 Flash Preview: 280B-param multi-router MoE activating only 14B per token — xiaohu · 2026-11-11
• ERNIE 5.5 recreates the Pelican test: 117-second animation drawn frame-by-frame in Python — xiaohu · 2026-11-11

Models channel →