OPT-BENCH-NP: leaderboard

The NP-hard half of OPT-BENCH (2026): an LLM agent iteratively improves solutions to ten classical NP-hard problems (graph coloring, knapsack, TSP and others) over 20 steps using validation feedback. The score is the expert-gap closure times 100, where 0 is the given initial solution and 100 the human expert heuristic; higher is better.

Metric: Expert-gap closure (EG) at 20 optimization steps with history (OPT-Agent): per problem, (current metric minus initial solution metric) divided by (human expert heuristic metric minus initial metric), averaged over the 10 NP-hard problems (graph coloring, Hamiltonian cycle, knapsack, maximum clique, maximum set, meeting schedule, minimum cut, set cover, subset sum, TSP), times 100; 0 is the initial solution and 100 the expert heuristic, and a run may fall below the one or exceed the other (a signed, unbounded scale); higher is better. Source: arxiv.org. Saturation forecast: Around November 2026. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1DeepSeek V3.1Thinking79#511
2Qwen 3 235B A22BThinking76#628
3O3 Mini73#546
4Qwen 3 30B A3B69#1156
5Claude 3.7 Sonnet (20250219)66#421
6Qwen 3 8B63#1873
7Grok 362#611
8Qwen 3 32B60#972
9Qwen 3 235B Instruct57#591
10GPT-4.153#505
11Claude 3.5 Sonnet (20241022)52#648
12Qwen 2.5 32B Instruct52#1247
13Qwen 2.5 72B Instruct47#1117
14GPT-4o (2024-08-06)47#743
15Qwen 2.5 14B Instruct47#1775

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=opt-bench-np · How It Works · Data refreshed daily, snapshot 2026-11-21.