ReliableMath - Prudence: leaderboard
The Prudence track of ReliableMath (arXiv) — whether models correctly abstain on unsolvable problems instead of hallucinating answers; score (%).
Metric: Score (%). Source: huggingface.co. Saturation forecast: Not forecast. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 1.5 | #336 |
| 2 | O3 Mini (2025-01-31) | 0.6 | #240 |
| 3 | DeepSeek R1 | 0.4 | #251 |
| 4 | Qwen 3 14B | 0.2 | #537 |
| 5 | DeepSeek V3 | 0.1 | #321 |
| 6 | Qwen 3 235B A22B | 0.1 | #313 |
| 7 | DeepSeek R1 Distill Qwen 32B | 0.1 | #655 |
| 8 | Gemini 2.5 Flash (Preview 04-17) | 0.1 | #237 |
| 9 | Gemini 2.5 Flash (Preview 04-17) (Thinking) | 0.1 | #237 (Gemini 2.5 Flash (Preview 04-17)) |
| 10 | Qwen 3 32B | 0 | #443 |
| 11 | DeepSeek R1 0528 | 0 | #222 |
| 12 | Claude Sonnet 4 (20250514) | 0 | #213 |
| 13 | DeepSeek R1 Distill Qwen 14B | 0 | #844 |
| 14 | DeepSeek-R1-Distill-Qwen-7B | 0 | #1415 |
| 15 | Claude Sonnet 4 (Thinking) | 0 | #199 (Claude Sonnet 4) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=reliablemath-prudence · How It Works · Data refreshed daily, snapshot 2026-10-20.