Neuron Kernel Generation on NKIGEN-BENCH 150-task balanced subset
20.7Pass RateSFT
Evaluation Results
| Method | Links | |
|---|---|---|
| SFTEvaluation Mode=NKIAgent (10 turns)2026.07 | 20.7 | |
| Base (Qwen3-30B-A3B)Evaluation Mode=NKIAgent (10 turns)2026.07 | 20 | |
| SFTEvaluation Mode=Single-Shot2026.07 | 19.3 | |
| GRPOEvaluation Mode=NKIAgent (10 turns)2026.07 | 17.3 | |
| GRPOEvaluation Mode=Single-Shot2026.07 | 15.3 | |
| Base (Qwen3-30B-A3B)Evaluation Mode=Single-Shot2026.07 | 14 |