OpenRef - Negative Rejection: leaderboard

Rejection of absent targets on OpenRef (2026), an open-world referring expression benchmark: for expressions whose object is not in the image, N3R rewards outputting no box and penalizes confident hallucinated boxes. Scored 0-100; higher is better.

Metric: N3R (0-100), negative relative rejection reliability: mean over expressions whose target is absent of the product of (1 - confidence) over the hallucinated boxes, average over five categories, zero-shot box grounding, each model prompted in its own grounding format; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2 VL 7B90.9#2275
2Qwen 3 VL 8B84.6#917
3GLM-4.6V-Flash78.7#1040
4LLaVA-OneVision-1.5-8B75.1#1864
5MiniCPM-V-4.573.6#1441
6OpenRef Mistral-3 8B (checkpoint unspecified)70.9
7Keye-VL-1.5-8B70.3#1720
8InternVL3.5-8B51#1604
9Qwen 3 VL 2B38.1#2134
10Qwen 3 VL 4B37.9#1151
11Qwen 2.5 VL 32B37.9#941
12MiMo-VL-7B-RL20.7#1155
13Qwen 2.5 VL 7B14.1#1862

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=openref-negative-rejection · How It Works · Data refreshed daily, snapshot 2026-11-21.