AutoRO-Bench: leaderboard
Robust-optimization reformulation on AutoRO-Bench (Chen et al., 2026): from a LaTeX LP or MILP with box, ellipsoidal or polyhedral uncertainty, the model derives the deterministic robust counterpart, and a fixed DeepSeek-V4-Pro coder turns it into solver code. Accuracy in percent on the 64 in-distribution Random instances, an instance counting when the solver optimum matches the ground truth within tolerance; zero-shot chain of thought, mean of three runs, higher is better.
Metric: Accuracy (%; share of the 64 in-distribution Random robust-optimization instances whose robust counterpart, turned into solver code by a fixed DeepSeek-V4-Pro coder, reaches the ground-truth optimal objective within tolerance; zero-shot CoT, mean of three runs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 90.6 | #121 |
| 2 | DeepSeek V4 Flash (Max) | 90.6 | #254 (DeepSeek V4 Flash) |
| 3 | DeepSeek V4 Flash (High) | 87.5 | #254 (DeepSeek V4 Flash) |
| 4 | Qwen 3.6 Plus | 81.2 | #217 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=autoro-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.