UnpredictaBench - RealWorld: leaderboard

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. Saturation forecast: Around 2036. 20 models tracked.

Top models

#ModelScore
1Llama 3.2 1B Instruct59.09
2DeepSeek V3.2 (Non-reasoning)36.67
3Llama 3.1 70B Instruct33.33
4Nemotron 3 Nano 30B A3B (Non-reasoning)33.33
5Qwen 3.5 2B (Non-reasoning)31.03
6Llama 3.1 8B Instruct30
7Qwen 3 32B (Non-reasoning)16.67
8GPT-4o Mini10
9Mistral Large 310
10Phi-3.5-mini-instruct10
11GPT-4o6.67
12Ministral-3-3B-Instruct-25126.67
13GPT-5.4 (Non-reasoning)6.67
14Qwen 3.5 35B A3B (Non-reasoning)3.57
15Claude Sonnet 4.63.33

Interactive version: theaggregate.ai/benchmark?slug=unpredictabench-realworld · How It Works · Data refreshed daily, snapshot 2026-09-29.