PerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning

Shaoxuan Li, Zhixuan Zhao, Hanze Deng, Zirun Ma, Shulin Tian, Zuyan Liu, Yushi Hu, Haoning Wu, Yuhao Dong, Benlin Liu, Ziwei Liu, Ranjay Krishna · 2026

Read PerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning by Shaoxuan Li, Zhixuan Zhao, Hanze Deng, Zirun Ma, Shulin Tian, Zuyan Liu, Yushi Hu, Haoning Wu, Yuhao Dong, Benlin Liu, Ziwei Liu, Ranjay Krishna (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-26653.txt.

Description

Li and co-authors introduce PerceptionComp, a video benchmark built so no single moment can answer its questions. Each item needs several pieces of visual evidence spread across time, joined by conjunctive or sequential logic. The 1,114 questions cover 279 videos, from city walks and home tours to games and extreme sports, all marked by hand. Humans score near chance, at 18.97%, when rewatching is banned. The best model tested, Gemini-3-Flash, reaches only 45.96%, far below its scores on older benchmarks.

Catalog record

CREATORShaoxuan Li, Zhixuan Zhao, Hanze Deng, Zirun Ma, Shulin Tian, Zuyan Liu, Yushi Hu, Haoning Wu, Yuhao Dong, Benlin Liu, Ziwei Liu, Ranjay Krishna
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52345
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.26653

Browse the full library · About the Tower of Babel