HyperGVL - Understanding: leaderboard

HyperGVL (2026) tests vision-language models on hypergraph understanding: counting vertices and hyperedges, degree- and order-specified counting and neighbor queries, each problem posed in 35 textual and visual representations. Average accuracy (%) over the six understanding tasks; higher is better.

Metric: Average accuracy (%) over the six understanding tasks (vertex, hyperedge, degree- and order-specified counting, neighbors, order-filtered neighbors), 7,000 questions per task (200 problems on synthetic and real-world hypergraphs, each posed in 35 textual, visual and combined hypergraph representations), accuracy (%) averaged over the representations, temperature 0.8, top-p 0.95; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash89.72#175
2Gemma 3 12B63.53#2059
3Qwen 3 VL 8B59.8#917
4InternVL3-8B44.96#1782
5Qwen 2.5 VL 7B Instruct42.39#1902
6GLM-4V-9B34.38#3142
7InternVL2.5-8B27.2#2370
8Molmo-7B-D-092418.6#7207
9DeepSeek-VL215.77#2927
10LLaVA-NeXT-7B11.44#5032
11LLaVA-v1.5-7B9.05#7630

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

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