Zelong Sun, Jun Wang, Kaicheng Yang, Tiancheng Gu, Ziyong Feng, Zhiwu Lu
4 min
Abstract
Unified multimodal retrieval aims to identify candidates that satisfy complex user intent expressed through heterogeneous inputs. Although Large Vision-Language Model (LVLM)-based retrievers are efficient and scalable, directly encoding raw multimodal inputs often misses fine-grained discriminative cues, leading to confusion among semantically similar candidates. Recent methods mitigate this limitation by generating Chain-of-Thought (CoT) rationales to enrich the query representation. However, such reasoning is typically derived from the query alone: it explains what the query describes, but not what the retriever misunderstands. We argue that effective retrieval reasoning should instead be conditioned on retrieval feedback. Based on this insight, we introduce UniME-R1, an embedder-adviser framework that learns to reason over initially retrieved candidates and generate Retrieval-Centric Chain-of-Thought (RC-CoT). The adviser analyzes candidates individually to identify the discriminative cues confused by the embedder. If the target appears in the initial top-k set, UniME-R1 directly reranks the candidates; otherwise, it generates RC-CoT to refine the retrieval direction and performs full-corpus re-retrieval with a dual-mode embedder. To train the framework, we mine hard negatives to simulate realistic retrieval failures, jointly optimize direct retrieval and RC-CoT-augmented retrieval, and align the adviser with retrieval outcomes through supervised learning and retrieval-oriented reinforcement learning. Extensive experiments on MMEB-V2 and a diverse set of general multimodal retrieval benchmarks demonstrate that UniME-R1 consistently improves retrieval performance over strong baselines.
Sam: You mentioned it also uses something called a "dual-mode embedder." What does that mean in practice?
Alex: So, when a search system processes a query or an image or a video clip, it converts that information into a kind of numerical fingerprint—a compact representation the computer can compare and rank. The dual-mode embedder means the system learns two different versions of that fingerprint: one optimized for a fast, broad scan across many results, and another for a deeper, more careful comparison when it needs to distinguish subtle differences.
Sam: Why does it need two modes? Why not just one good fingerprint?
Alex: Because the tasks are genuinely different. A fast scan needs to be efficient—you're comparing against thousands of results. A deep comparison needs to be precise—you're trying to tell apart things that look nearly identical. Training both in the same model is part of what makes UniME-R1 work across images, video, and text without a separate system for each.
Sam: And the system is trained on examples that are almost right but not quite?
Alex: Yes. The researchers train it using what they call "hard negatives"—results that look very similar to the correct answer but are actually wrong. It forces the model to learn the fine-grained differences that separate a near-miss from a genuine match. And then on top of that, they use a reinforcement learning technique that rewards the model specifically when it correctly identifies a failure and produces a fix that works.
Sam: So it's not just memorizing correct answers from a textbook. It's learning to judge whether its own output is actually good.
Alex: Precisely. That self-correcting loop is what the paper argues makes the system more reliable in practice—not just on the specific tasks it was trained on, but in the messier, more varied searches that real users actually perform. It's a meaningful step toward AI that doesn't just guess, but understands when and why it has guessed wrong.
Sam: That's a genuinely useful distinction. Thanks for walking through it.
Alex: Thanks for listening to ResearchPod.