跳转至

文章背景与核心概要

生成式检索（Generative Retrieval）通过将表征学习和搜索统一为一个单一的序列到序列（sequence-to-sequence）生成任务，近年来受到了广泛关注。然而，将这一范式扩展到跨模态检索时，会引入一个重大的结构性挑战：简洁的文本查询与稠密的视觉候选项之间存在严重的信息不对称。

这种不匹配导致使用前缀树约束束搜索（trie-constrained beam search）的标准自回归解码器遭受强制幻觉（forced hallucination）的困扰。由于模型在未能猜出原始查询中缺失的细粒度细节时会受到惩罚，导致不相关的候选项经常劫持顶部排名。

为了克服这一问题，作者推出了 WIDE（Wildcard Inference with Dynamic Expansion，具有动态扩展的通配符推理），它由三个核心组件构成： 1. 自适应熵阈值化（AET）： 离线校准层特定的不确定性边界。 2. 感知不对称的通配符解码（AWD）： 在生成过程中检测语义盲区，并输出通配符而不是强制的确定性标识符，从而在不产生对数概率惩罚的情况下扩大搜索空间。 3. 盲区重排（BSR）： 使用结合了离散生成置信度与连续语义相似度的混合评分机制，对扩展后的候选池进行评估。

在 M-BEIR 基准测试上的实验表明，WIDE 的性能优于最先进的生成式检索技术，同时有效地抑制了强制幻觉，并保持了紧凑的索引结构。

WIDE: Wildcard Inference with Dynamic Expansion for Cross-Modal Generative Retrieval

Summary

Generative retrieval has gained significant traction by unifying representation learning and search into a single sequence-to-sequence generation task. However, extending this paradigm to cross-modal retrieval introduces a major structural challenge: information asymmetry between concise text queries and dense visual candidates.

This mismatch causes standard autoregressive decoders using trie-constrained beam search to suffer from forced hallucination. Because models are penalized when failing to guess fine-grained details missing from the original query, irrelevant candidates frequently hijack top rankings.

To overcome this, the authors introduce WIDE (Wildcard Inference with Dynamic Expansion), which consists of three key components: 1. Adaptive Entropy Thresholding (AET): Calibrates layer-specific uncertainty boundaries offline. 2. Asymmetry-aware Wildcard Decoding (AWD): Detects semantic blind spots during generation and emits wildcards instead of forced deterministic identifiers, expanding the search space without log-probability penalties. 3. Blind-Spot Re-ranking (BSR): Evaluates the expanded candidate pool using a hybrid scoring mechanism combining discrete generation confidence with continuous semantic similarity.

Experiments on the M-BEIR benchmark show that WIDE outperforms state-of-the-art generative retrieval techniques while effectively suppressing forced hallucinations and keeping index structures compact.

Paper Metadata

  • arXiv Identifier: arXiv:2609.03554 [cs.CV]
  • Authors: Teng Guo, Xin Wang, Jiayou Xu, Keying Zhou, Jifeng Shen, Haoxin Ruan
  • Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Secondary Subjects: Artificial Intelligence (cs.AI)
  • Accepted Venue: 34th ACM International Conference on Multimedia (ACM MM 2026)
  • Submission Date: September 3, 2026

Abstract

生成式检索通过将表征学习和搜索统一为一个单一的序列到序列生成任务，已经取得了显著的成功。然而，将这一范式扩展到跨模态检索时，暴露出一个由不同模态之间固有的信息不对称所引发的关键挑战，例如简洁文本查询与稠密视觉候选之间的鸿沟。这种结构上的不匹配导致自回归解码器在使用标准的前缀树约束束搜索生成标识符时，会遭受强制幻觉的困扰：当模型未能猜出查询中缺少的细粒度细节时，会受到严厉惩罚，从而允许不相关的候选者劫持顶部排名。为了解决这个问题，我们提出了具有动态扩展的通配符推理（WIDE）。WIDE 采用自适应熵阈值化（AET）离线校准层特定的不确定性边界。在解码生成阶段，感知不对称的通配符解码（AWD）检测语义盲区并输出通配符，而不是强制的确定性标识符，从而在不产生对数概率惩罚的情况下动态扩大搜索空间。最后，盲区重排（BSR）使用结合了离散生成置信度与连续语义相似度的混合评分机制来评估扩展后的候选池。在 M-BEIR 基准测试上的大量实验表明，WIDE 优于最先进的生成式检索方法，在保持紧凑索引结构的同时，有效抑制了强制幻觉。

Generative retrieval has demonstrated significant success by unifying representation learning and search into a single sequence-to-sequence generation task. However, extending this paradigm to cross-modal retrieval reveals a critical challenge arising from the inherent information asymmetry across different modalities, such as the gap between concise text queries and dense visual candidates. This structural mismatch causes the autoregressive decoder to suffer from forced hallucination when generating identifiers via standard trie-constrained beam search, where the model is severely penalized for failing to guess fine-grained details absent from the query, allowing irrelevant candidates to hijack top rankings. To address this issue, we propose Wildcard Inference with Dynamic Expansion (WIDE). WIDE employs Adaptive Entropy Thresholding (AET) to calibrate layer-specific uncertainty boundaries offline. During the decoding generation phase, Asymmetry-aware Wildcard Decoding (AWD) detects semantic blind spots and emits wildcards instead of forced deterministic identifiers, dynamically expanding the search space without incurring log-probability penalties. Finally, Blind-Spot Re-ranking (BSR) evaluates the expanded candidate pool using a hybrid scoring mechanism that combines discrete generation confidence with continuous semantic similarity. Extensive experiments on the M-BEIR benchmark demonstrate that WIDE outperforms state-of-the-art generative retrieval methods, effectively suppressing forced hallucination while maintaining compact index structures.

Full-Trust & Resources / 全文与资源