OpenRef - Polysemy: leaderboard

The polysemy split of OpenRef (2026), an open-world referring expression benchmark: expressions built on polysemous words, to be grounded with bounding boxes in the image. The score is F1 over predicted and ground-truth boxes in percent; higher is better.

Metric: Grounding F1 (%) on expressions built on polysemous words, 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 3 VL 8B71.4#927
2Qwen 3 VL 4B68.6#1163
3GLM-4.6V-Flash68.2#1049
4MiMo-VL-7B-RL65.5#1160
5Qwen 2.5 VL 7B64.5#1878
6Keye-VL-1.5-8B64.4#1729
7Qwen 3 VL 2B63.5#2156
8Qwen 2.5 VL 32B61.4#944
9LLaVA-OneVision-1.5-8B59.1#1876
10MiniCPM-V-4.554.9#1451
11InternVL3.5-8B52.1#1618
12Qwen 2 VL 7B46.8#2300
13OpenRef Mistral-3 8B (checkpoint unspecified)4.6

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=openref-polysemy · How It Works · Data refreshed daily, snapshot 2026-11-22.