Saved in:
Bibliographic Details
Main Authors: Li, Shengzhi, Chen, Jiarun, Sharma, Karun, Su, Jiaqi, Pei, Shichao
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.23407
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914507651022848
author Li, Shengzhi
Chen, Jiarun
Sharma, Karun
Su, Jiaqi
Pei, Shichao
author_facet Li, Shengzhi
Chen, Jiarun
Sharma, Karun
Su, Jiaqi
Pei, Shichao
contents Large vision-language models (VLMs) can recognize \textit{what} happens in video but fail to count \textit{how many} times. We introduce \textbf{PushupBench}, 446 long-form clips (avg. 36.7s) for evaluating repetition counting. The best frontier model achieves 42.1\% exact accuracy; open-source 4B models score $\sim$6\%, matching supervised baselines. We show that accuracy alone misleads -- weaker models exploit the modal count rather than reason temporally. Fine-tuning on counting with 1k samples transfers to general video understanding: MVBench (+2.15), PerceptionTest (+1.88), TVBench (+4.54), suggesting counting is a proxy for broader temporal reasoning.PushupBench incorporated in \texttt{lmms-eval} (https://github.com/EvolvingLMMs-Lab/lmms-eval/pull/1262) and hosted on (pushupbench.com/)
format Preprint
id arxiv_https___arxiv_org_abs_2604_23407
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle PushupBench: Your VLM is not good at counting pushups
Li, Shengzhi
Chen, Jiarun
Sharma, Karun
Su, Jiaqi
Pei, Shichao
Computer Vision and Pattern Recognition
Artificial Intelligence
Large vision-language models (VLMs) can recognize \textit{what} happens in video but fail to count \textit{how many} times. We introduce \textbf{PushupBench}, 446 long-form clips (avg. 36.7s) for evaluating repetition counting. The best frontier model achieves 42.1\% exact accuracy; open-source 4B models score $\sim$6\%, matching supervised baselines. We show that accuracy alone misleads -- weaker models exploit the modal count rather than reason temporally. Fine-tuning on counting with 1k samples transfers to general video understanding: MVBench (+2.15), PerceptionTest (+1.88), TVBench (+4.54), suggesting counting is a proxy for broader temporal reasoning.PushupBench incorporated in \texttt{lmms-eval} (https://github.com/EvolvingLMMs-Lab/lmms-eval/pull/1262) and hosted on (pushupbench.com/)
title PushupBench: Your VLM is not good at counting pushups
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2604.23407