KnowHal - Negative Questions: leaderboard

Metric: Accuracy (%; mean over the entity, attribute, relation and knowledge dimensions on questions carrying a misleading premise the model must reject; 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 2032. 14 models tracked.

Top models

#ModelScore
1Qwen 2.5 VL 72B Instruct57
2Qwen 3 VL 32B Instruct56.3
3Gemini 3 Flash (Preview)49.3
4Kimi K2.541.8
5GPT-5.240.4
6Seed 2.0 Mini37.1
7InternVL3.5-8B33.3
8GLM-4.6V31.2
9GPT-5.6 Sol30.6

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