Title: Training-Free Composed Video Retrieval via Visual Representation-Guided Video-LLM Reasoning

Abstract:

The emergence of large vision-language models has significantly broadened the scope of video retrieval, moving beyond basic text queries to more adaptable scenarios. In these advanced settings, users can define their desired outcomes by combining visual examples with textual directives. The CVPR 2026 Reason-Aware Composed Video Retrieval Challenge specifically requires systems to identify a target video based on a reference clip alongside a modification instruction. To tackle this challenge, we propose a framework for Training-Free Composed Video Retrieval that leverages Visual Representation-Guided Video-LLM Reasoning.

Our approach initially employs frozen DINOv3 models to generate a concise pool of visually relevant candidates. Subsequently, large vision-language models are utilized to assess whether each candidate meets the specified modification criteria. To further enhance performance, a reasoning-based refinement process is applied to the top-ranked results, aiming to optimize the primary prediction. Operating without any training phase, our system secures a Recall@1 of 48.78 and a Recall@5 of 51.48 on the test set. Looking ahead, retrieval precision could be further enhanced by integrating more powerful video-LLMs and achieving deeper synchronization between visual representations and linguistic reasoning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC