UnpredictaBench - Code: leaderboard
UnpredictaBench (2026) tests whether LLMs can sample outcomes that follow a target probability distribution rather than collapsing onto one plausible answer. This board is KS@100 on the 238 Code problems, whose distributions are induced by stochastic programs: 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 238 Code problems (distributions induced by stochastic programs) 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. Saturation forecast: Not forecast. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 29.41 | #741 |
| 2 | nemotron-3-nano-30B-a3B (Non-reasoning) | 18.49 | #1875 (nemotron-3-nano-30B-a3B) |
| 3 | Ministral-3-3B-Instruct-2512 | 17.23 | #2579 |
| 4 | DeepSeek V3.2 (Non-reasoning) | 14.29 | #406 (DeepSeek V3.2) |
| 5 | Qwen 3.5 2B (Non-reasoning) | 12.61 | #4810 (Qwen 3.5 2B) |
| 6 | GPT-4o Mini | 10.5 | #1623 |
| 7 | GPT-5.4 (Non-reasoning) | 10.5 | #123 (GPT-5.4) |
| 8 | Llama 3.1 70B Instruct | 9.66 | #1546 |
| 9 | Qwen 3 32B (Non-reasoning) | 5.46 | #960 (Qwen 3 32B) |
| 10 | OLMo 3 7B Instruct | 5.46 | #6860 |
| 11 | Grok 4.1 Fast (Non-reasoning) | 5.46 | #430 (Grok 4.1 Fast) |
| 12 | Claude Sonnet 4.6 | 5.04 | #161 |
| 13 | Mistral Large 3 | 5.04 | #817 |
| 14 | Llama 3.1 8B Instruct | 4.62 | #4246 |
| 15 | Llama 3.2 1B Instruct | 4.2 | #9047 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=unpredictabench-code · How It Works · Data refreshed daily, snapshot 2026-11-13.