UnpredictaBench - RealWorld: leaderboard
KS@100 on the 30 RealWorld problems, natural-language scenarios of random processes such as race conditions or multi-outcome decisions, from UnpredictaBench (2026), a benchmark of sampling from target distributions: the percent of problems for which 100 samples from the model, drawn at temperature 1.0 with reasoning off, are not rejected by a two-sample Kolmogorov-Smirnov test against ground-truth samples at p below 0.0001. Higher is better.
Metric: KS@100 (%): share of the 30 RealWorld problems (natural-language scenarios of random processes) for which 100 samples from the model (100 independent calls at temperature 1.0, reasoning effort none, up to 5 retries for a parseable answer) are not rejected by a two-sample Kolmogorov-Smirnov test against ground-truth samples at p < 0.0001; higher is better. Source: arxiv.org. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.2 1B Instruct | 59.09 | #9047 |
| 2 | DeepSeek V3.2 (Non-reasoning) | 36.67 | #406 (DeepSeek V3.2) |
| 3 | Llama 3.1 70B Instruct | 33.33 | #1546 |
| 4 | nemotron-3-nano-30B-a3B (Non-reasoning) | 33.33 | #1875 (nemotron-3-nano-30B-a3B) |
| 5 | Qwen 3.5 2B (Non-reasoning) | 31.03 | #4810 (Qwen 3.5 2B) |
| 6 | Llama 3.1 8B Instruct | 30 | #4246 |
| 7 | Qwen 3 32B (Non-reasoning) | 16.67 | #960 (Qwen 3 32B) |
| 8 | GPT-4o Mini | 10 | #1623 |
| 9 | Mistral Large 3 | 10 | #817 |
| 10 | Phi-3.5-mini-instruct | 10 | #5733 |
| 11 | GPT-4o | 6.67 | #741 |
| 12 | Ministral-3-3B-Instruct-2512 | 6.67 | #2579 |
| 13 | GPT-5.4 (Non-reasoning) | 6.67 | #123 (GPT-5.4) |
| 14 | Qwen 3.5 35B A3B (Non-reasoning) | 3.57 | #482 (Qwen 3.5 35B A3B) |
| 15 | Claude Sonnet 4.6 | 3.33 | #161 |
Interactive version: theaggregate.ai/benchmark?slug=unpredictabench-realworld · How It Works · Data refreshed daily, snapshot 2026-11-13.