CoDyControlBench: leaderboard
CoDyControlBench (2026) asks LLMs to write and tune feedback controllers for 132 simulated dynamical systems of one to six degrees of freedom, with feedback from each closed-loop simulation. A design succeeds when every channel meets the steady-state band and overshoot limits; the score is the percentage of successful designs under the Fixed-3 protocol (three iterations, best design counts). Higher is better.
Metric: Design success rate (%; Fixed-3 protocol: the model always runs three design iterations with feedback and the best successful controller counts; 132 controller-design configurations (1-6 DoF, LTI/LTV/nonlinear systems, coupling, damping, SMC- and PID-family controllers), mean of three independent runs; a design succeeds when every response channel meets the steady-state band and overshoot limits). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (High) | 94.8 | #49 (GPT-5.5) |
| 2 | Gemini 3.1 Pro (Preview) | 90.4 | #94 |
| 3 | GLM-5.2 | 82.3 | #138 |
| 4 | Claude Opus 4.8 | 58 | #56 |
| 5 | DeepSeek V4 Pro (High) | 55.1 | #188 (DeepSeek V4 Pro) |
| 6 | Qwen 3.5 397B A17B | 50 | #246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=codycontrolbench · How It Works · Data refreshed daily, snapshot 2026-11-13.