Code Generation on LCB V6 (Throughput and Accuracy)
75.9AccuracyRSA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RSAModel 1=GPT-OSS-20B, Model 2=GPT-OSS-120B, Obs. M1 share=0%, GPU split (G2:G1)=12:02026.04 | 75.9 | 5.66 | 1 | — | |
| SQUEEZE EVOLVE (p=10)Model 1=GPT-OSS-20B, Model 2=GPT-OSS-120B, Obs. M1 share=87.1%, GPU split (G2:G1)=4:82026.04 | 75.6 | 14.3 | 2.53 | — | |
| SQUEEZE EVOLVE (p=0)Model 1=GPT-OSS-20B, Model 2=GPT-OSS-120B, Obs. M1 share=100%, GPU split (G2:G1)=0:122026.04 | 73.3 | 19.02 | 3.36 | — | |
| DREBase Model=Qwen3-30B-A3B-Thinking-25072026.06 | 64.86 | — | — | 13,977 | |
| S-GRPOBase Model=Qwen3-30B-A3B-Thinking-25072026.06 | 64.79 | — | — | 17,028 | |
| Raw ModelBase Model=Qwen3-30B-A3B-Thinking-25072026.06 | 64.64 | — | — | 15,559 | |
| GRPO + LPBase Model=Qwen3-30B-A3B-Thinking-25072026.06 | 63.93 | — | — | 15,494 | |
| GRPOBase Model=Qwen3-30B-A3B-Thinking-25072026.06 | 63.43 | — | — | 15,413 | |
| RSAModel 1=Qwen3-30B-A3B-I, Model 2=Qwen3-235B-A22B-I, Obs. M1 share=0%, GPU split (G2:G1)=16:02026.04 | 59.1 | 3.83 | 1 | — | |
| SQUEEZE EVOLVE (p=10)Model 1=Qwen3-30B-A3B-I, Model 2=Qwen3-235B-A22B-I, Obs. M1 share=87.5%, GPU split (G2:G1)=8:82026.04 | 55.9 | 15.07 | 3.93 | — | |
| SQUEEZE EVOLVE (p=0)Model 1=Qwen3-30B-A3B-I, Model 2=Qwen3-235B-A22B-I, Obs. M1 share=100%, GPU split (G2:G1)=0:162026.04 | 55.3 | 31.42 | 8.2 | — | |
| Raw ModelBase Model=Qwen3-4B-Thinking-25072026.06 | 54.26 | — | — | 17,722 | |
| S-GRPOBase Model=Qwen3-4B-Thinking-25072026.06 | 54.14 | — | — | 18,460 | |
| DREBase Model=Qwen3-4B-Thinking-25072026.06 | 53.79 | — | — | 14,630 | |
| GRPOBase Model=Qwen3-4B-Thinking-25072026.06 | 53.64 | — | — | 16,333 | |
| GRPO + LPBase Model=Qwen3-4B-Thinking-25072026.06 | 53.57 | — | — | 16,663 | |
| GRPOBase Model=Qwen3-8B2026.06 | 49.5 | — | — | 12,197 | |
| DREBase Model=Qwen3-8B2026.06 | 48.93 | — | — | 11,210 | |
| S-GRPOBase Model=Qwen3-8B2026.06 | 48.86 | — | — | 12,293 | |
| GRPO + LPBase Model=Qwen3-8B2026.06 | 48.64 | — | — | 11,995 | |
| Raw ModelBase Model=Qwen3-8B2026.06 | 47.93 | — | — | 12,177 |