Commonsense Reasoning on Commonsense Reasoning (Comprehensive Suite)
75.1BoolQ AccuracyEcho-DoRA
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Echo-DoRAModel=LLaMA3-8B2026.05 | 75.1 | 89 | 81.9 | 88.5 | 81.6 | 91.3 | 86.8 | 97.1 | — | 86.4 | |
| SAMoRABackbone=Llama3.1-8B, TP (%)=0.152026.04 | 74.89 | 90.37 | 83.32 | 88.95 | 86.35 | 94.7 | 89.8 | 95.97 | 84.85 | 87.64 | |
| SAMoRABackbone=Qwen3-8B, TP (%)=0.152026.04 | 74.68 | 92 | 83.78 | 88.95 | 92.58 | 97.94 | 91.8 | 96.01 | 87.31 | 91.71 | |
| DoRAModel=LLaMA3-8B2026.05 | 74.6 | 89.3 | 79.9 | 85.6 | 80.4 | 90.5 | 85.8 | 95.5 | — | 85.2 | |
| MoOREBackbone=Llama3.1-8B, TP (%)=0.772026.04 | 74.49 | 88.63 | 82.99 | 87.74 | 79.95 | 88.8 | 86.2 | 95.48 | 84.6 | 85.43 | |
| MTL-LoRABackbone=Llama3.1-8B, TP (%)=0.162026.04 | 74.34 | 89.9 | 82.95 | 88.08 | 84.55 | 93.81 | 88.2 | 95.15 | 83.94 | 86.77 | |
| HydraLoRABackbone=Llama3.1-8B, TP (%)=0.172026.04 | 74.31 | 90.15 | 82.49 | 88.47 | 84.06 | 92.18 | 87.8 | 93.18 | 83.81 | 86.27 | |
| MoELoRABackbone=Qwen3-8B, TP (%)=0.562026.04 | 73.9 | 91.18 | 81.47 | 83.1 | 92.49 | 97.34 | 89.6 | 92.3 | 84.43 | 87.31 | |
| LoRABackbone=Qwen3-8B, TP (%)=0.742026.04 | 73.8 | 91.45 | 83 | 88.39 | 92.4 | 97.6 | 90.2 | 94.6 | 86.32 | 88.64 | |
| MoOREBackbone=Qwen3-8B, TP (%)=0.842026.04 | 73.6 | 91.26 | 80.8 | 86.55 | 90.1 | 93.3 | 90.2 | 94.09 | 86.56 | 90.28 | |
| MTL-LoRABackbone=Qwen3-8B, TP (%)=0.162026.04 | 73.51 | 91.13 | 82.08 | 88.87 | 92.15 | 97.55 | 91.4 | 95.47 | 86.08 | 90.98 | |
| Echo-DoRAModel=LLaMA2-7B2026.05 | 73.5 | 84.7 | 81.6 | 85.6 | 74.2 | 88.1 | 87.4 | 94.4 | — | 83.7 | |
| HydraLoRABackbone=Qwen3-8B, TP (%)=0.162026.04 | 73.14 | 90.69 | 83.21 | 87.92 | 92.9 | 97.47 | 89.4 | 94.6 | 87.01 | 90.33 | |
| ExpertCondenserModel=Qwen1.5-MoE, Model Size=14B, Post-train method=ExpertCondenser (Ours)2026.04 | 72.1 | 84.9 | 75.6 | 79.8 | 78.1 | 88.5 | 84.4 | 81.6 | — | 80.6 | |
| MultiLoRABackbone=Qwen3-8B, TP (%)=0.292026.04 | 71.89 | 89.88 | 81.83 | 83.89 | 92.15 | 97.6 | 90.6 | 93.07 | 85.74 | 87.64 | |
| DoRAModel=LLaMA2-7B2026.05 | 71.8 | 83.7 | 76 | 82.6 | 68.2 | 83.7 | 82.4 | 89.1 | — | 79.7 | |
| MultiLoRABackbone=Llama3.1-8B, TP (%)=0.262026.04 | 70.95 | 80.81 | 80.91 | 82.15 | 71.7 | 86.12 | 80.6 | 94.01 | 80.34 | 80.84 | |
| DenseMixerModel=Qwen1.5-MoE, Model Size=14B, Post-train method=DenseMixer2026.04 | 70.8 | 85.7 | 74.6 | 78.9 | 74.8 | 82.6 | 77.8 | 75.8 | — | 77.6 | |
| LoRABackbone=Llama3.1-8B, TP (%)=2.092026.04 | 70.43 | 82.97 | 76 | 71.11 | 77.56 | 85.77 | 81.6 | 93 | 77.4 | 79.54 | |
| ESFTModel=Qwen1.5-MoE, Model Size=14B, Post-train method=ESFT2026.04 | 69.7 | 85.3 | 75.4 | 74.2 | 71.8 | 84 | 75 | 78.2 | — | 76.7 | |
| DoRAModel=LLaMA-7B2026.05 | 69.7 | 83.4 | 78.6 | 81 | 66.2 | 81.9 | 79.2 | 87.2 | — | 78.4 | |
| Echo-DoRAModel=LLaMA-7B2026.05 | 69.6 | 82.6 | 81.2 | 83.6 | 70.7 | 86.1 | 83.2 | 93.4 | — | 81.3 | |
| SFTModel=Qwen1.5-MoE, Model Size=14B, Post-train method=SFT2026.04 | 68.8 | 84.7 | 74.5 | 75.6 | 72.8 | 84.6 | 76.4 | 76.8 | — | 76.8 | |
| ExpertCondenserModel=OLMoE, Model Size=7B, Post-train method=ExpertCondenser (Ours)2026.04 | 67.7 | 71.4 | 69.7 | 75.8 | 71 | 76.8 | 73.6 | 69.8 | — | 71.9 | |
| ESFTModel=OLMoE, Model Size=7B, Post-train method=ESFT2026.04 | 63.5 | 63 | 58.9 | 62.8 | 63.8 | 74.8 | 63.4 | 64.7 | — | 64.4 | |
| DenseMixerModel=OLMoE, Model Size=7B, Post-train method=DenseMixer2026.04 | 62.8 | 68.7 | 65.3 | 73.5 | 63.5 | 81.3 | 71.3 | 71.6 | — | 69.8 | |
| SFTModel=OLMoE, Model Size=7B, Post-train method=SFT2026.04 | 62.5 | 65.8 | 62.8 | 71.4 | 63.7 | 78.4 | 70.6 | 70.7 | — | 68.2 |