UMER Unifies Multimodal Embedding and Ranking for Retrieval

Universal multimodal retrieval systems aim to handle diverse instruction-aware tasks, requiring both efficient large-scale matching and precise semantic understanding. Existing methods often rely on Multimodal Large Language Models (MLLMs) to derive representations or use Chain-of-Thought (CoT) reasoning to enhance embeddings. However, these CoT methods typically reason about queries and candidates in isolation, failing to explicitly highlight differences between positive and confusing negative examples. This research introduces UMER (Unified Multimodal Embedding and Ranking), a novel framework designed for universal multimodal retrieval. UMER departs from item-wise reflection by employing Pair-Aware Discriminative Reasoning. This approach directly compares query-candidate pairs to pinpoint instruction-relevant matching and discrepancy evidence, providing a more nuanced understanding. UMER jointly learns contrastive embeddings for efficient global matching and discriminative ranking for explicit pairwise relevance judgments, all within a single MLLM. A mutual distillation strategy further enhances the transfer of reliable pairwise preferences between the embedding and ranking functions. On the MMEB-V2 benchmark, UMER achieved state-of-the-art performance under comparable settings, while also offering adjustable inference budgets.