FrontierSWE: FrogsGame Post-Training (Mean@5)
Code · 2026-04-16
Post-train Qwen3-8B to solve FrogsGame boards through tool use. FrontierSWE Mean@5 is the average solve score across up to five independent long-horizon agent runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 34.0 |
| Grok 4.5 | 3.4 |
| Opus 4.8 | 3.0 |
| GLM-5.2 | 2.4 |
| GPT-5.4 | 2.0 |
| Composer 2.5 | 1.4 |
| GPT-5.5 | 1.4 |
| Opus 4.6 | 1.0 |
| Opus 4.7 | 1.0 |
| DeepSeek-V4-Pro | 0.4 |
| GLM-5.1 | 0.4 |
| Kimi K2.6 | 0.4 |
| Gemini 3.1 Pro Preview | 0.0 |
| Kimi K2.5 | 0.0 |
| Qwen3.6 Plus (2026-04-02) | 0.0 |