Pinterest Uses VLM for Search Relevance Eval, Cutting Labeling Turnaround by 20x

_reachsumit · x · 2026-08-04

Pinterest deploys a VLM-based relevance evaluation pipeline for Search, fine-tuning Qwen3-VL on human-annotated query-Pin pairs, reducing labeling turnaround by over 20x. The method validates alignment with human annotations, improves evaluation efficiency, and enables expanding query sets and optimizing sampling, significantly reducing MDEs in online experiments. Paper at RecSys'26 Industry track.

Original post →

More from Apps

Apps channel →