NSMQ Riddles (Real-Time Proxy): leaderboard
Real-time proxy setting of NSMQ Riddles (Boateng et al., Kwame AI and ETH Zurich, 2026) on the 156 riddles of the 2019 contests: clues arrive seven words at a time, as in a live quiz, and after each chunk the model gives an answer and says whether it is sure; its answer is taken at the first sure response or after the last clue. Exact match in percent against the answer or its listed variants; higher is better.
Metric: Exact match (%; 156 riddles from 2019, clues revealed seven words at a time and the answer taken when the model says it is sure, or after the last clue). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 75.64 | #121 (GPT-5.4) |
| 2 | Gemini 3.1 Pro (Preview) (High) | 73.72 | #91 (Gemini 3.1 Pro (Preview)) |
| 3 | GPT-5.4 (Low) | 73.72 | #121 (GPT-5.4) |
| 4 | Gemini 3.1 Pro (Preview) (Low) | 71.15 | #91 (Gemini 3.1 Pro (Preview)) |
| 5 | Claude Opus 4.6 (High) | 51.28 | #100 (Claude Opus 4.6) |
| 6 | GPT-OSS-120B (High) | 46.79 | #699 (GPT-OSS-120B) |
| 7 | GPT-OSS-120B (Low) | 46.15 | #699 (GPT-OSS-120B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=nsmq-riddles-real-time-proxy · How It Works · Data refreshed daily, snapshot 2026-11-12.