# CaptionQA outperforms VideoQA on 10 of 12 models for long egocentric videos > Source: > Published: 2026-09-17 07:56:05.395008+00:00 [arXiv](https://arxiv.org/abs/2609.17688) ### CaptionQA outperforms VideoQA on 10 of 12 models for long egocentric videos Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Textual captions of egocentric video segmented into 30-to-60-second windows outperform direct video querying on videos longer than 20 minutes, delivering a 3.22-point accuracy gain that increases by up to 5.3 points when paired with a retrieve-and-verify architecture. For engineers building wearable AI agents, this means you can bypass expensive visual-token costs and long-context vision failures by storing and querying a lightweight text-caption index of video history instead of feeding raw video frames into vision-language models. On egocentric videos longer than 20 minutes, caption-based memory beat direct video QA in 10/12 models with 30-second caption windows and kept a 3.22-point mean accuracy gain in matched-frame controls. For production wearable/agent systems, precomputing dense text captions as episodic memory is a practical way to reduce visual-token pressure and improve long-horizon recall, especially when paired with retrieve-and-verify for another accuracy boost of up to 5.3 points.