AIAny
Icon for item

Learning from Failures: Retrieval-Centric CoT via Hard Negatives for Unified Multimodal Retrieval

Generates retrieval-centric Chain-of-Thought (RC-CoT) over initially retrieved candidates to improve unified multimodal retrieval via reranking or full-corpus re-retrieval with a dual-mode embedder. Trains an embedder–adviser framework (UniME-R1) using mined hard negatives, supervised learning, and retrieval-oriented reinforcement learning.

Introduction

Most multimodal retrievers encode raw inputs directly and miss fine-grained cues that distinguish semantically similar candidates. The paper’s core insight is that effective retrieval reasoning must be conditioned on retrieval feedback — i.e., what the retriever got wrong — not only on the query itself. Conditioning reasoning on initially retrieved candidates lets an adviser detect confused cues and either rerank or guide a re-retrieval to recover the target.

Key Findings
  • Embedder–adviser architecture (UniME-R1): the embedder produces an initial ranked list; the adviser inspects top-k candidates and generates Retrieval-Centric Chain-of-Thought (RC-CoT) that highlights discriminative cues the embedder missed. If the target is in top-k, the system reranks; otherwise it uses RC-CoT to steer a dual-mode embedder for full-corpus re-retrieval. This design separates fast retrieval from reasoning-driven refinement.
  • Training with realistic failures: the authors mine hard negatives to simulate retrieval mistakes, jointly optimize direct retrieval and RC-CoT-augmented retrieval, and align the adviser to retrieval outcomes via supervised objectives plus retrieval-oriented reinforcement learning. This makes the adviser produce rationales that actually improve downstream retrieval utility, not just plausible explanations.
  • Empirical gains: evaluated on MMEB-V2 and a diverse suite of multimodal retrieval benchmarks, UniME-R1 consistently improves recall and reranking metrics over strong LVLM-based baselines, particularly on cases requiring fine-grained discrimination.
Who it's for and trade-offs

Great fit if you need higher accuracy on compositional or fine-grained multimodal retrieval tasks and can afford a two-stage pipeline where an initial fast embedder is augmented by an adviser for difficult queries. Look elsewhere if you require single-shot, lowest-latency retrieval with no opportunity for re-ranking or full-corpus re-retrieval, or if compute and latency budgets cannot accommodate a dual-mode embedder plus reasoning model.

Where it fits

UniME-R1 targets the middle ground between pure embedding-based retrieval (fast but brittle to fine-grained confusions) and heavy reranking approaches: it keeps an efficient first-stage embedder while using retrieved-candidate-conditioned reasoning to decide when and how to refine results, making it suitable for systems that can trade some latency for notable accuracy gains.

Information

  • Websitearxiv.org
  • AuthorsZelong Sun, Jun Wang, Kaicheng Yang, Tiancheng Gu, Ziyong Feng, Zhiwu Lu
  • Published date2026/08/06

More Items

Evaluates vision-language model judges on computer-using agent (CUA) trajectories to measure verifier reliability. Provides OSReward-Hard and OSReward-Multi challenge sets, the OS-Shepherd-100K reasoning-annotated corpus, and trained OS-Shepherd reward models that match commercial judges at ~30–60× lower cost.

Generates large-scale, explorable 3D open-world scenes from open-ended text prompts, producing editable instance-level assets and a consistent global terrain. Uses agentic planning to convert text into region/terrain/asset specifications and a coarse-to-fine pipeline for terrain construction, mesh reconstruction, and render-based refinement.

Systematically studies how language and vision interact during unified multimodal pretraining, identifies mechanisms that enable modality synergy versus competition, demonstrates the benefit of early joint training, and derives efficient pretraining recipes validated at scale.