GraphVerse - Common Neighbor - Accuracy: leaderboard
GraphVerse (arXiv 2608.06769v1, Table 1): multimodal models solve for the common neighbors of two authors in a co-authorship graph, reading a rendered graph image altered by recoloring, patch perturbation, composition or attention focusing. Answer accuracy on the 100-item testmini split for this task, in percent; higher is better.
Metric: Answer accuracy (%; testmini). Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 89 | #103 |
| 2 | GPT-5.2 | 87 | #115 |
| 3 | Gemini 3 Pro | 86 | #82 |
| 4 | Qwen-VL-Max | 76 | #475 |
| 5 | GLM-4.6V | 66 | #330 |
| 6 | Kimi K2.5 (Thinking) | 66 | #152 (Kimi K2.5) |
| 7 | O1 | 64 | #190 |
| 8 | Grok 4 Fast | 47 | #266 |
| 9 | Qwen 3.5 35B A3B | 19 | #268 |
| 10 | QVQ-72B-Preview | 18 | #458 |
| 11 | Qwen 3.5 4B | 15 | #514 |
| 12 | InternVL3-8B | 14 | #624 |
| 13 | InternVL3-38B | 7 | #413 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=graphverse-common-neighbor-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.