---
格式版本: 2
标题: "UMER: Unifying Embedding and Ranking via Pair-Aware Discriminative Reasoning for Universal Multimodal Retrieval"
原文链接: "https://arxiv.org/abs/2608.18504"
发布日期: "2026-08-19"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Wed, 19 Aug 2026 03:51:39 UTC (2,985 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-20T14:38:07+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-20T14:35:51+08:00"
入库时间: "2026-08-20T06:38:07.841Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=Meta&searchtype=all"
匹配关键词:
  - "performance"
  - "AI"
相关厂家:
  - "Meta"
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 4
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "该arXiv论文为多模态检索的UMER框架，属人工智能算法研究，未涉及超节点、AI Rack、机柜级基础设施、供电散热互连或量产落地等主题，与项目关注范围完全无关。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-20T14:38:44+08:00"
AI主题相关性: 0
AI来源权威性: 2
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 2
AI摘要: "UMER提出统一多模态嵌入与排序框架，用pair-aware discriminative reasoning比较查询-候选对来区分正向与难负样本，并在同一多模态大模型中联合学习对比嵌入和判别排序。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:18:43.847Z"
采集批次: "2026年8月20日14点19分32秒"
采集批次ID: "20260820-141932-079"
去重键: "https://arxiv.org/abs/2608.18504"
---

## Computer Science > Artificial Intelligence

## Title:UMER: Unifying Embedding and Ranking via Pair-Aware Discriminative Reasoning for Universal Multimodal Retrieval

Authors:[Libiao Chen](https://arxiv.org/search/cs?searchtype=author&query=Chen,+L), [Xiyang Liu](https://arxiv.org/search/cs?searchtype=author&query=Liu,+X), [Yanheng Wei](https://arxiv.org/search/cs?searchtype=author&query=Wei,+Y), [Tao Wang](https://arxiv.org/search/cs?searchtype=author&query=Wang,+T), [Zhenyu Tang](https://arxiv.org/search/cs?searchtype=author&query=Tang,+Z)

[View PDF](https://arxiv.org/pdf/2608.18504) [HTML (experimental)](https://arxiv.org/html/2608.18504v1)

> Abstract:Universal multimodal retrieval aims to support diverse instruction-aware retrieval tasks, demanding both efficient corpus-scale matching and fine-grained semantic reasoning. Recent MLLM-based embedding methods typically derive representations from hidden states, while Chain-of-Thought (CoT) reasoning is emerging as a promising strategy for embedding enhancement by encoding intermediate semantic evidence into the representation space. However, existing CoT methods typically use item-wise reasoning over queries and candidates in isolation, providing no explicit evidence to distinguish a positive from a semantically confusable hard negative. Moreover, contrastive embeddings capture global similarity but struggle with meta-tasks requiring answer verification, category judgment or fine-grained reasoning. In this paper, we propose UMER, a Unified Multimodal Embedding and Ranking framework for universal multimodal retrieval. UMER replaces item-wise reflection with Pair-Aware Discriminative Reasoning, which compares query--candidate pairs to identify instruction-relevant matching and discrepancy evidence. UMER jointly learns contrastive embeddings for efficient global matching and discriminative ranking for explicit pairwise relevance judgment within a single MLLM. A complementary mutual distillation strategy further transfers reliable pairwise preferences between the embedding and ranking functions. On the MMEB-V2 benchmark, UMER achieves state-of-the-art performance under comparable experimental settings while supporting budget-adjustable inference.

| Subjects: | Artificial Intelligence (cs.AI) |
| --- | --- |
| Cite as: | [arXiv:2608.18504](https://arxiv.org/abs/2608.18504) \[cs.AI\] |
|  | (or [arXiv:2608.18504v1](https://arxiv.org/abs/2608.18504v1) \[cs.AI\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.18504](https://doi.org/10.48550/arXiv.2608.18504) |

## Submission history

From: Xiyang Liu \[[view email](https://arxiv.org/show-email/e84fa60b/2608.18504)\]  
**\[v1\]** Wed, 19 Aug 2026 03:51:39 UTC (2,985 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.18504) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))