Robotic task planning and code generation on Advanced
90Success RateCorrective GSCE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Corrective GSCE2026.04 | 90 | 97.8 | |
| Corrective GSCE2026.04 | 81.7 | 96.3 | |
| GSCE2026.04 | 75 | 91.5 | |
| Ro-SLMtraining tasks=Complex Task Only2026.04 | 75 | 83.8 | |
| GSCE2026.04 | 75 | 94.8 | |
| Ro-SLMvariant=without GRPO Optimization2026.04 | 75 | 86.4 | |
| Ro-SLM2026.04 | 75 | 87 | |
| Ro-SLMGRPO optimization=none2026.04 | 70 | 83.1 | |
| Ro-SLM2026.04 | 70 | 83.7 | |
| Ro-SLMvariant=without Augmentation2026.04 | 70 | 81.3 | |
| Ro-SLMvariant=Complex Task Only2026.04 | 70 | 83 | |
| Ro-SLMaugmentation=none2026.04 | 55 | 74.8 | |
| Ro-SLMvariant=with Comments2026.04 | 55 | 75.2 | |
| Ro-SLMcomments=retained2026.04 | 50 | 76.2 | |
| Llama-3.1-8B2026.04 | 5 | 9 | |
| Ro-SLMtraining tasks=Simple Task Only2026.04 | 0 | 1.9 | |
| Llama-3.1-8B2026.04 | 0 | 8.4 | |
| Ro-SLMvariant=Simple Task Only2026.04 | 0 | 3.6 |