Work

Beyond Query Bias: Candidate-Aware Iterative Refinement for Zero-Shot Composed Image Retrieval

Year: 2026

Type: conference-paper

Abstract: Zero-Shot Composed Image Retrieval (ZS-CIR) retrieves target images using a reference image and modification text without taskspecific training.Existing methods typically rely on MLLMs to generate query vectors with pre-trained models like CLIP.However, those constructed queries suffer from inherent cognitive bias due to unknown candidate distribution.We propose CoRR, a training-free framework that reframes ZS-CIR as a self-correcting process through bias-aware query refinement.CoRR uses retrieved results as feedback to perceive the candidate distribution.With carefully designed CoT prompting, the MLLM inspects the retrieved candidates to identify intent misalignments in the query and then corrects them via Historical Query Fusion.We also introduce Retrieval-Driven Caption Optimization to provide context-aligned examples, reducing phrasing and style mismatches.Experiments on public benchmarks show that CoRR significantly outperforms other SOTA methods.

Source: Findings of the Association for Computational Linguistics: ACL (Findings of ACL)

DOI: 10.18653/v1/2026.findings-acl.1120

Authors: Nan Sun, Jing Tang, Lei Sun, Rui Chen, Yuxing Lu, Xiangxiang Chu, Hefei Ling, Yujun Cai

Institutions: Alibaba Group (China), Huazhong University of Science and Technology, Peking University, The University of Queensland

Language: English

Topic: Image Retrieval and Classification Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

OpenAlex ID: W7166855952

FWCI:

Cites:

Cited by:

Related to: 10