Research map / Vision-language and multimodal / Multimodal LLMs

Beyond Global Similarity: Multi-Conditional Retrieval for Fine-Grained Cross-Modal Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MCMR, a 10,400-product benchmark for multi-condition, cross-modal (image+text) product retrieval, evaluating MLLM-based retrievers and pointwise MLLM rerankers.

Problem

Existing multimodal retrieval benchmarks test coarse or single-condition alignment and rarely require multiple interdependent constraints grounded in both images and long-form text.

Approach

A benchmark whose queries combine complementary visual-only and text-only attributes so that a candidate is relevant only if all conditions across both modalities are satisfied, built with a Qwen-based generation and verification pipeline plus human checks; it also benchmarks MLLM pointwise rerankers.

Key results

• Best first-stage Recall@1 reaches only 26.57% (CORAL) under fused candidates, while LLaVE achieves the highest Recall@100 at 78.64%.
• lychee-reranker-mm as pointwise reranker on LLaVE top-50 reaches nDCG@1 of 82.51, versus 26.57 for the best first-stage retriever's NDCG@1.
• Removing image-derived constraints from text-only queries drops R@10 by roughly 20-30 points relative to fused queries.

Surprising findings

• Parameter count does not predict reranking ability; architecture and vision-text grounding matter more, per the authors.
• LLaVE nearly collapses on image-only candidates (R@1 drops from 24.99 to 0.90).

Limitations

• Pointwise reranking is computationally expensive and does not scale to large corpora.
• Queries are synthetic (LLM-generated) and only validated on a 100-sample human study.
• Evaluation is limited to product retrieval in four clothing/jewelry/shoes/furniture domains.

Experimental setup

Datasets: Amazon Reviews 2023

Benchmarks: MCMR

Metrics: Recall@K nDCG@K MRR@10

Baselines: GME-Qwen2-VL-7B LLaVE-7B VLM2Vec LamRA-Ret-Qwen2.5-VL-7B CORAL MM-EMBED

Builds on: MERIT MultiConIR Amazon Reviews 2023 FashionIQ CIRR

Closest papers by meaning

• ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge (NeurIPS 2026 · similarity 0.91)
• MRMR: A Realistic and Expert-Level Multidisciplinary Benchmark for Reasoning-Intensive Multimodal Retrieval (ICLR 2026 · similarity 0.91)
• RMIR: A Benchmark Dataset for Reasoning-Intensive Multimodal Image Retrieval (CVPR 2026 · similarity 0.91)
• F2RVLM: Boosting Fine-grained Fragment Retrieval for Multi-Modal Long-form Dialogue with Vision Language Model (AAAI 2026 · similarity 0.90)
• MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS (ICLR 2025 · similarity 0.90)
• MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models (ICLR 2025 · similarity 0.89)