NPHardEval: leaderboard
Evaluates LLMs on 9 NP-hard computational problems (TSP, graph coloring, knapsack, etc.) spanning P, NP-complete, and NP-hard complexity classes.
Metric: Average Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo | 37.74 | #404 |
| 2 | Claude 2 | 26.42 | #761 |
| 3 | GPT-3.5 Turbo | 25.86 | #865 |
| 4 | Qwen-14B-Chat | 22.28 | #1169 |
| 5 | Yi Chat 34B | 18.93 | #976 |
| 6 | Mistral Instruct v0.1 7B | 18.22 | #1454 |
| 7 | Phi-2 | 9.21 | #1511 |
| 8 | vicuna-13B-v1.3 | 7.56 | #1447 |
| 9 | Phi-1.5 | 0.22 | #1549 |
| 10 | mpt-30B Instruct | 0.02 | #1470 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nphardeval · How It Works · Data refreshed daily, snapshot 2026-10-20.