Commonsense Reasoning on Commonsense Reasoning
85AccuracyMultiplex CoT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Multiplex CoTReasoning Method=Multiplex Chain of Thought2025.01 | 85 | 9 | 12 | — | |
| Self-DistillationBackbone=Llama3-8B2026.05 | 81.9 | — | — | — | |
| DASDBackbone=Llama3-8B2026.05 | 81.6 | — | — | — | |
| Hint-decodingBackbone=Llama3-8B2026.05 | 81.4 | — | — | — | |
| Mask-PPLBackbone=Llama3-8B2026.05 | 81 | — | — | — | |
| BaseBackbone=Llama3-8B2026.05 | 79.9 | — | — | — | |
| CoTReasoning Method=Standard Chain of Thought2025.01 | 78 | — | — | — | |
| Baseline (full model)Sparsity=0%, Backbone=Phi-4-14B2025.05 | 74.38 | — | — | — | |
| WINASparsity=25%, Backbone=Phi-4-14B2025.05 | 74.25 | — | — | — | |
| TEALSparsity=25%, Backbone=Phi-4-14B2025.05 | 74.12 | — | — | — | |
| WINASparsity=40%, Backbone=Phi-4-14B2025.05 | 73.76 | — | — | — | |
| WINASparsity=50%, Backbone=Phi-4-14B2025.05 | 73.51 | — | — | — | |
| TEALSparsity=40%, Backbone=Phi-4-14B2025.05 | 73.49 | — | — | — | |
| TEALSparsity=50%, Backbone=Phi-4-14B2025.05 | 72.73 | — | — | — | |
| DensePruning Ratio=0%, Backbone=Phi-4-14B, Zero-shot=true2025.12 | 72.66 | — | — | — | |
| QLoRABit=4, Model=Qwen2.5-7B-Ins, Fine-tuning=true2026.05 | 71.59 | — | — | — | |
| WINASparsity=65%, Backbone=Phi-4-14B2025.05 | 71.58 | — | — | — | |
| DensePruning Ratio=0%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 71.4 | — | — | — | |
| Token FilteringPruning Ratio=20%, Backbone=Phi-4-14B, Zero-shot=true2025.12 | 70.52 | — | — | — | |
| DensePruning Ratio=0%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 70.39 | — | — | — | |
| TEALSparsity=65%, Backbone=Phi-4-14B2025.05 | 69.93 | — | — | — | |
| FP16Bit=16, Model=Qwen2.5-7B-Ins, Fine-tuning=true2026.05 | 69.84 | — | — | — | |
| DenseKV cache budget%=100%, TOPS(token/s)=4.7, Memory(MB)=51531.9, Backbone=LLaMA-2-13B2025.12 | 69.51 | — | — | — | |
| ProjQBit=4, Model=LLaMA2-7B, Fine-tuning=true2026.05 | 68.7 | — | — | — | |
| QLoRABit=4, Model=LLaMA2-7B, Fine-tuning=true2026.05 | 68.63 | — | — | — | |
| FP16Bit=16, Model=LLaMA2-7B, Fine-tuning=true2026.05 | 68.07 | — | — | — | |
| Token FilteringPruning Ratio=20%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 67.94 | — | — | — | |
| LoftQBit=4, Model=LLaMA2-7B, Fine-tuning=true2026.05 | 67.21 | — | — | — | |
| H2OKV cache budget%=50%, TOPS(token/s)=8.2 (+76%), Memory(MB)=38444.1 (-25.4%), Backbone=LLaMA-2-13B2025.12 | 66.99 | — | — | — | |
| DensePruning Ratio=0%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 66.88 | — | — | — | |
| ProjQBit=4, Model=Qwen2.5-7B-Ins, Fine-tuning=true2026.05 | 66.78 | — | — | — | |
| Token FilteringKV cache budget%=50%, TOPS(token/s)=8.8 (+87%), Memory(MB)=38421.3 (-25.5%), Backbone=LLaMA-2-13B2025.12 | 65.9 | — | — | — | |
| Token FilteringPruning Ratio=20%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 65.82 | — | — | — | |
| LoftQBit=4, Model=Qwen2.5-7B-Ins, Fine-tuning=true2026.05 | 64.56 | — | — | — | |
| SlimGPT w/oPruning Ratio=20%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 64.27 | — | — | — | |
| Token FilteringPruning Ratio=20%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 63.61 | — | — | — | |
| PPPruning Ratio=20%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 63.34 | — | — | — | |
| FLAPPruning Ratio=20%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 63.1 | — | — | — | |
| SlimGPT w/oPruning Ratio=20%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 62.76 | — | — | — | |
| Token FilteringPruning Ratio=50%, Backbone=Phi-4-14B, Zero-shot=true2025.12 | 59.87 | — | — | — | |
| Token FilteringPruning Ratio=50%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 59.31 | — | — | — | |
| PPPruning Ratio=20%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 58.66 | — | — | — | |
| FLAPPruning Ratio=20%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 57.72 | — | — | — | |
| FLAPPruning Ratio=20%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 55.75 | — | — | — | |
| SlimGPT w/oPruning Ratio=20%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 55.51 | — | — | — | |
| Token FilteringPruning Ratio=50%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 53.13 | — | — | — | |
| SlimGPT w/oPruning Ratio=50%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 51.45 | — | — | — | |
| PPPruning Ratio=50%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 50.59 | — | — | — | |
| Token FilteringPruning Ratio=50%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 50.37 | — | — | — | |
| SlimGPT w/oPruning Ratio=50%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 47.68 | — | — | — | |
| SlimGPT w/oPruning Ratio=20%, Backbone=Phi-4-14B, Zero-shot=true2025.12 | 47.61 | — | — | — | |
| FLAPPruning Ratio=50%, Backbone=LLaMA-2-7B, Zero-shot=true2025.12 | 46.97 | — | — | — | |
| FLAPPruning Ratio=50%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 44.57 | — | — | — | |
| SlimGPT w/oPruning Ratio=50%, Backbone=Mistral-7B, Zero-shot=true2025.12 | 43.2 | — | — | — | |
| FLAPPruning Ratio=50%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 42.61 | — | — | — | |
| PPPruning Ratio=50%, Backbone=LLaMA-3-8B, Zero-shot=true2025.12 | 41.54 | — | — | — | |
| SlimGPT w/oPruning Ratio=50%, Backbone=Phi-4-14B, Zero-shot=true2025.12 | 32.53 | — | — | — | |
| MyGO Multiplex CoT (MCoT)Reflection Type=Single-reflection2026.01 | — | 0.85 | 12 | 9 | |
| MyGO Poly-Reflective CoT (PR-CoT)Reflection Type=Multi-perspective reflection2026.01 | — | 0.87 | 14 | 11 | |
| Traditional CoT2026.01 | — | 0.78 | — | — |