Matching with Deliberation: Test-Time Evolutionary Hierarchical Multi-Agents for Zero-Shot Compositional Image Retrieval
Researchers have introduced a novel framework called PDF for zero-shot compositional image retrieval. This hierarchical multi-agent system aims to overcome limitations in existing methods by incorporating experience self-evolution and Test-Time Scaling Law (TTS). The framework dynamically routes perception signals and employs a training-free reasoning policy distillation with a tournament-style TTS strategy for fine-grained reasoning, achieving state-of-the-art results on benchmark datasets. AI
IMPACT Introduces a novel approach to zero-shot image retrieval, potentially improving performance in applications requiring fine-grained understanding of image modifications.