Key Info
A new benchmark called "World Models' Last Exam in Physics" tested 8 state-of-the-art video world models on 40 controlled tasks, with the top score reaching only 57.76/100. MiniMax H3 claims the best open-source performance, nearly matching closed-source models.
Highlights
- Top score across 8 SoTA video world models was just 57.76/100, showing physics remains a hard challenge
- Free-fall videos passed consistency screening but averaged only 26.61/100 composite score
- Benchmark consists of 40 controlled physics tasks for evaluating world models
- MiniMax H3 positions itself as the leading open-source video world model in physics understanding