JOR-Bench - NL4OPT: leaderboard

Metric: Accuracy (%; the 245 NL4OPT linear programming word problems, Japanese translation with human post-editing; the model writes a standalone Python program in a fenced block, run with OR-Tools available (300-second and 32 GB limits), correct when the printed value matches the reference within 1e-6; higher is better). Source: arxiv.org. Saturation forecast: Around December 2026. 7 models tracked.

Top models

#ModelScore
1Qwen 3 8B78.4
2Qwen 3 32B70.6
3Qwen3-Swallow-8B-RL-v0.246.5
4Qwen3-Swallow-32B-RL-v0.242.9
5NVIDIA-Nemotron-Nano-9B-v2-Japanese4.5

Interactive version: theaggregate.ai/benchmark?slug=jor-bench-nl4opt · How It Works · Data refreshed daily, snapshot 2026-09-29.