GPTNT - Simulator VQA: leaderboard
Offline visual question answering from GPTNT (Parekh et al., TMLR 2026): 640 questions about rendered Keep Talking and Nobody Explodes scenes that ask a model to read text, name colours or count elements, or to recognise that a question is ambiguous or refers to something absent. Accuracy in percent pooled over all questions; higher is better.
Metric: Accuracy (%; 640 questions about rendered game scenes: reading, colour, counting, ambiguity and hallucination detection, pooled over items). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (Minimal) | 70 | #141 (Gemini 3 Flash (Preview)) |
| 2 | Claude Sonnet 4.6 (Non-reasoning) | 63 | #158 (Claude Sonnet 4.6) |
| 3 | GPT-5.2 (2025-12-11) (Non-reasoning) | 58.9 | |
| 4 | Qwen 3.5 27B (Non-reasoning) | 56.9 | #358 (Qwen 3.5 27B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gptnt-simulator-vqa · How It Works · Data refreshed daily, snapshot 2026-11-12.