MMLU Redux
87.47AccuracySSA-LLM-8B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SSA-LLM-8BInference Mode=think2026.03 | 87.47 | — | |
| Qwen3-8BInference Mode=think2026.03 | 87.42 | — | |
| SSA-LLM-8BInference Mode=no-think2026.03 | 83.47 | — | |
| Qwen3-8BInference Mode=no-think2026.03 | 81.12 | — | |
| Cache-to-CacheSharer=Llama3.2-1B, Receiver=Qwen3-0.6B2025.10 | 44.42 | 0.5 | |
| Cache-to-CacheSharer=Qwen3-4B-Base, Receiver=Qwen3-0.6B2025.10 | 43.95 | 0.45 | |
| Text-to-TextSharer=Qwen3-4B-Base, Receiver=Qwen3-0.6B2025.10 | 43.87 | 7.54 | |
| Text-to-TextSharer=Llama3.2-1B, Receiver=Qwen3-0.6B2025.10 | 43.32 | 0.75 | |
| Cache-to-CacheSharer=Qwen2.5-0.5B, Receiver=Qwen3-0.6B2025.10 | 42.92 | 0.4 | |
| Text-to-TextSharer=Qwen2.5-0.5B, Receiver=Qwen3-0.6B2025.10 | 41.03 | 1.52 | |
| SharerModel=Qwen2.5-0.5B2025.10 | 38.42 | 0.34 | |
| RoutingSharer=Qwen2.5-0.5B, Receiver=Qwen3-0.6B2025.10 | 35.58 | 0.27 | |
| ReceiverModel=Qwen3-0.6B2025.10 | 35.53 | 0.29 | |
| RoutingSharer=Llama3.2-1B, Receiver=Qwen3-0.6B2025.10 | 33.38 | 0.18 | |
| SharerModel=Llama3.2-1B2025.10 | 32.3 | 0.06 | |
| RoutingSharer=Qwen3-4B-Base, Receiver=Qwen3-0.6B2025.10 | 16.39 | 0.28 | |
| SharerModel=Qwen3-4B-Base2025.10 | 1.03 | 2.06 |