ReCoG CUB-Hard - ReCoG Weighted Accuracy: leaderboard

ReCoG CUB-Hard - ReCoG Weighted Accuracy. The ReCoG paper evaluates frozen MiniCPM-o-2.6 on CUB-Hard, a subset of 20 bird species. Retrieval uses CLIP ViT-L and one reference image per concept, without negative database examples. Baseline and ReCoG contexts are separate conditions. Positive-query accuracy, negative-query accuracy, and the source weighted average are separate metrics; baseline and ReCoG negative queries are evaluated from their top-1 retrieval results. The LLaVA-1.5 comparison is held because the paper does not identify its exact size or checkpoint. Reported metric: Accuracy (%).

Metric: Accuracy (%). Source: arxiv.org. 1 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=recog-cub-hard-recog-weighted-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-25.