Real-world image search queries are often multimodal and compositional, such as "find this shirt in pink," which requires retaining an entity, modifying an attribute, and ignoring context. However, existing re-rankers either compress multifaceted relevance into an opaque embedding or rely on free-form chain-of-thought that can omit or hallucinate fine-grained constraints.
Drawing on rubric- and checklist-based evaluation from NLP, researchers at Hugging Face propose EviRank, which recasts multimodal image re-ranking as a semantic constraint satisfaction problem. EviRank parses any query—text-only, image-only, or composed—into a unified evidence package with typed criteria across six semantic slots (e.g., entities, attributes, relations), each labelled required, forbidden, or ignorable.
Re-ranking then reduces to evidence-conditioned verification, combining deterministic rubric scoring and evidence-grounded listwise comparison in a single training-free procedure. The explicit evidence can also serve as structured supervision for optionally distilling a lightweight student model.
Across five benchmarks spanning text-to-image, image-to-image, and composed image retrieval, EviRank achieves state-of-the-art performance. The distilled student preserves over 90% of the teacher's capability at substantially lower cost.