PushupBench: Your VLM is not good at counting pushups

TL;DR

PushupBench introduces a new dataset to evaluate vision-language models' ability to count repetitions in videos, revealing current models' limitations and the potential for counting to improve broader temporal reasoning.

Contribution

The paper presents PushupBench, a dataset for counting in videos, and demonstrates how fine-tuning on counting tasks enhances general video understanding and temporal reasoning.

Findings

Best model achieves 42.1% accuracy, open-source models score around 6%.

We show that accuracy alone can be misleading, as weaker models exploit modal counts.

Fine-tuning on counting improves performance on other temporal reasoning benchmarks.

Abstract

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/)