EviRank: Structured Relevance Evidence for Multimodal Image Re-ranking
EviRank turns image re-ranking into explicit constraint checking instead of embedding-only matching.
The paper parses text, image, or composed queries into typed evidence across six semantic slots, marking criteria as required, forbidden, or ignorable. It then verifies candidate images against that evidence with deterministic rubric scoring and evidence-grounded listwise comparison, without training a new reranker. The authors report state-of-the-art results across five retrieval benchmarks. A distilled student model keeps more than 90% of the teacher’s capability at lower cost. HF Daily Papers' note
The paper parses text, image, or composed queries into typed evidence across six semantic slots, marking criteria as required, forbidden, or ignorable. It then verifies candidate images against that evidence with deterministic rubric scoring and evidence-grounded listwise comparison, without training a new reranker. The authors report state-of-the-art results across five retrieval benchmarks. A distilled student model keeps more than 90% of the teacher’s capability at lower cost. HF Daily Papers' note
score 4