FrontierSWE: Optimizer Design (Mean@5)
Code · 2026-04-16
Design a single optimizer that beats tuned AdamW across diverse ML workloads. FrontierSWE Mean@5 is the average multiplicative improvement across up to five independent long-horizon agent runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Grok 4.5 | 2.7 |
| GLM-5.2 | 2.6 |
| Opus 4.7 | 2.3 |
| GPT-5.5 | 2.1 |
| Opus 4.8 | 2.0 |
| GPT-5.4 | 1.6 |
| Opus 4.6 | 1.5 |
| Composer 2.5 | 1.5 |
| Kimi K2.6 | 1.5 |
| GLM-5.1 | 1.3 |
| DeepSeek-V4-Pro | 1.2 |
| Gemini 3.1 Pro Preview | 1.1 |
| Kimi K2.5 | 1.1 |
| Qwen3.6 Plus (2026-04-02) | 0.9 |
| Claude Fable 5 | 0.8 |