KnowHal - Relation (Negative): leaderboard

Metric: Accuracy (%; relation-dimension negative (false-premise) questions; 1,800 entity-image samples in 10 domains with paired positive and false-premise negative questions; positive answers scored by standard VQA answer normalization and soft matching, negative answers by a GPT-4o-mini judge (option matching for relation questions); zero-shot). Source: arxiv.org. Saturation forecast: Around 2028. 14 models tracked.

Top models

#ModelScore
1Qwen 2.5 VL 72B Instruct73.9
2Gemini 3 Flash (Preview)71.9
3Qwen 3 VL 32B Instruct64.6
4GPT-5.263.7
5Kimi K2.557.1
6Seed 2.0 Mini55.5
7GPT-5.6 Sol52.2
8GLM-4.6V43.2
9InternVL3.5-8B38.3

Interactive version: theaggregate.ai/benchmark?slug=knowhal-relation-negative · How It Works · Data refreshed daily, snapshot 2026-09-29.