AI Paper Atlas

Research map / Vision-language and multimodal / Cross-modal learning

Robust Cross-Modal Retrieval via Generative Semantic Refinement and Exclusion-Guided Adaptation

Qin Yang, Xin Wei, Yanjia Li, Jiasun Feng, Mingrui Zhu, Nannan Wang, Xinbo Gao

ICML 2026 · Vision-language and multimodal › Cross-modal learning · approach: Probabilistic embeddings for cross-modal retrieval

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace raw query with LLM-rectified text directly (remove CGDF) (ablation)ReEx GSR · Flickr-C T2IRecall@1: 65.7 -> 63.2Hurt
add Confidence-Guided Dynamic Fusion to LLM rectification (ablation)ReEx GSR · Flickr-C T2IRecall@1: 63.2 -> 65.7Improved
add Exclusion-aware Negative Mining to proxy alignment (ablation)ReEx EPCL · Flickr-C T2IRecall@1: 64.3 -> 65.7Improved
use soft proxy over top-K candidates instead of hard pseudo-labels (k=1) (ablation)ReEx EPCL · Flickr-C T2IRecall@1: 65.5 -> 65.7Improved
add LLM rectification alone without fusion or EPCL (ablation)BLIP without adaptation · Flickr-C T2IRecall@1: 62.1 -> 62.5No clear effect
add GSR (LLM and CGDF) on top of EPCL (ablation)ReEx · Flickr-C T2IRecall@1: 63.3 -> 65.7Improved
use ReEx instead of TCR under query shiftTTA for BLIP · Flickr-C I2T and T2IRecall@1: 72.4 -> 74.8 (I2T); 63.4 -> 65.7 (T2I)Improved
use ReEx instead of TCR under query-gallery shiftTTA for BLIP · Query-Gallery Shift averageRecall@1: 66.4 -> 66.8No clear effect
use ReEx with BLIP-2 backboneTTA for BLIP-2 · Query shiftaverage Recall@1: 65.6 -> 67.1Improved
use larger LLM (Llama-3.2-1B) in GSR (ablation)ReEx · Flickr-C T2IRecall@1: 65.7 -> 67.0 with 1.6x inference timeImproved
use smaller LLM (SmolLM2-360M) in GSR (ablation)ReEx · Flickr-C T2IRecall@1: 65.7 -> 65.1Hurt
vary neighborhood size K from 1 to 9 (ablation)ReEx EPCL · Flickr-C T2IRecall@1: variance 0.003No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy