GPTNT - Manual VQA Procedural Reasoning: leaderboard
Offline manual questions from GPTNT (Parekh et al., TMLR 2026): 499 questions that ask a model to apply the Keep Talking and Nobody Explodes manual's rules to a described bomb state and choose the Defuser's correct action, from single rule lookups to multi-step conditional rules. Accuracy in percent; higher is better.
Metric: Accuracy (%; 499 questions applying the manual's rules to a described bomb state to choose the Defuser's action). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 (Non-reasoning) | 67.1 | #158 (Claude Sonnet 4.6) |
| 2 | Gemini 3 Flash (Preview) (Minimal) | 64.1 | #141 (Gemini 3 Flash (Preview)) |
| 3 | GPT-5.2 (2025-12-11) (Non-reasoning) | 61.5 | |
| 4 | Qwen 3.5 27B (Non-reasoning) | 53.7 | #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-manual-vqa-procedural-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-12.