The authors built The Prestige, a video understanding benchmark drawn from professional magic performances, because magic forces observers to track long visual contexts, hold expectations, and explain surprising outcomes without relying on language or captions.
Magic-trick videos expose the reasoning gap between AI models and humans
In 2,177 open-ended questions about professional magic performances, the best model scored about half as well as people
Top University
Shuo Wen · Beixi Du · Edwin Meriaux · Junming Shi · Chloe Si · David Meger · +2 more
McGill University · Mila – Quebec AI Institute · Cornell University · Université Paris-Saclay · CentraleSupélec
Research Digest··2 min read
In one line
Magic-trick videos expose a large gap between vision-language models and people in tracking events, reasoning about causes, and resisting misleading questions.
Before you read this
Two ideas this paper builds onShuo Wen and colleagues built The Prestige, a benchmark of 2,177 open-ended questions grounded in professional magic performances, and tested 14 leading vision-language models against human baselines.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.