Original title: MARS: What Retrieval Signals Are Hidden in Multimodal Large Language Models for Text-Video Retrieval?
Uicheol Jung, Juyoung Hong, Geuntaek Lim, Yukyung Choi · September 2026
Finding the right video using a text search is tricky because many clips share similar scenes, actions, or timing. Many modern AI models tackle this by squishing an entire video or text prompt into just one single summary vector taken from the model's final layer. This can squeeze out crucial small details. To fix this, researchers created a new framework named MARS. Instead of one summary, MARS builds multiple representation slots by mixing information across different layers of the model. It then compares the matching text and video slots and adds up their scores. A special training goal also helps it tell tricky, look-alike videos apart. Across four benchmark tests, MARS achieved state-of-the-art results for retrieving videos.
Imagine you're trying to match a written description to one of many similar-looking photos. If you only glance at one feature, like color, you might pick the wrong one. But if you examine multiple details—shape, background, and texture—you're more likely to find the exact match. In this paper, MARS does something similar: it looks at various layers of the model to get different features for videos and text, so it can tell them apart even when they're confusing.
This helps anyone searching for videos online get more accurate results. It affects platforms like YouTube and research that relies on video retrieval.
The abstract does not specify limitations, only that MARS achieves state-of-the-art results on four text-video retrieval benchmarks.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
MARS: What Retrieval Signals Are Hidden in Multimodal Large Language Models for Text-Video Retrieval? — Uicheol Jung, Juyoung Hong, Geuntaek Lim et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.