List your company
ResearchFormerly WizWand

Grade School Math Word Problems on GSM8K

0.971AccuracyQwen-3.5 27B
-0.0333280.2274110.488150.748889Jan 8, 2024Jun 2, 2024Oct 26, 2024Mar 22, 2025Aug 15, 2025Jan 8, 2026Jun 4, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Qwen-3.5 27BConfig=Hybrid MoE, @32k speedup=0.5×2026.040.971
TDA-RCBase LLM=DeepSeek-V32026.030.944
TDA-RCBase LLM=GPT-4o-mini2026.030.942
TDA-RCBase LLM=Qwen-Turbo2026.030.937
Instruction InductionBase LLM=DeepSeek-V32026.030.934
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.050.933
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.050.932
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.050.93
Instruction InductionBase LLM=GPT-4o-mini2026.030.929
Role / Persona PromptingBase LLM=DeepSeek-V32026.030.928
BaseModel backbone=Qwen3-4B-It, Fine-tuning method=Base2026.050.926
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.050.926
Instruction InductionBase LLM=Qwen-Turbo2026.030.925
Role / Persona PromptingBase LLM=GPT-4o-mini2026.030.924
Prompt CanvasBase LLM=DeepSeek-V32026.030.924
Prompt CanvasBase LLM=GPT-4o-mini2026.030.921
Role / Persona PromptingBase LLM=Qwen-Turbo2026.030.92
HoTBase LLM=DeepSeek-V32026.030.92
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.050.919
BaseModel backbone=Qwen3-8B, Fine-tuning method=Base2026.050.918
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=02026.050.918
Prompt CanvasBase LLM=Qwen-Turbo2026.030.917
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.050.917
HoTBase LLM=GPT-4o-mini2026.030.916
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=02026.050.915
HoTBase LLM=Qwen-Turbo2026.030.914
OPSDModel backbone=Qwen3-8B, Fine-tuning method=OPSD2026.050.882
Analogical PromptingBase LLM=DeepSeek-V32026.030.876
Analogical PromptingBase LLM=GPT-4o-mini2026.030.872
Analogical PromptingBase LLM=Qwen-Turbo2026.030.87
OPSDModel backbone=Qwen3-4B-It, Fine-tuning method=OPSD2026.050.839
Qwen3-4B2026.020.8362
PretrainRLBackbone=Qwen3-4B2026.020.8097
BaseModel backbone=Qwen3-1.7B, Fine-tuning method=Base2026.050.804
CoDDDecoding Strategy=Entropy, Diffusion Steps=2562026.020.7475
Mixtral 8x7BActive Params=13B2024.010.744
DreamDecoding Strategy=Entropy, Diffusion Steps=2562026.020.7134
DreamDecoding Strategy=Margin, Diffusion Steps=2562026.020.7043
LLaMA 2 70BActive Params=70B2024.010.696
CoDDDecoding Strategy=Entropy, Diffusion Steps=1282026.020.6702
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.050.656
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.050.622
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.050.607
CoDDDecoding Strategy=Entropy, Diffusion Steps=642026.020.5641
DreamDecoding Strategy=Entropy, Diffusion Steps=1282026.020.5618
DreamDecoding Strategy=Margin, Diffusion Steps=1282026.020.561
GIFTBackbone Model=Dream-v0-7B-Base, Training Dataset=Tulu3-10k, Fine-tuning Protocol=Full-parameter Tuning, Evaluation Protocol=0-shot2025.090.552
SFTBackbone Model=Dream-v0-7B-Base, Training Dataset=Tulu3-10k, Fine-tuning Protocol=Full-parameter Tuning, Evaluation Protocol=0-shot2025.090.503
Mistral 7BActive Params=7B2024.010.5
YOCO (CLSA)Model Scale=4B, Attention=Cross-Layer Sparse2026.060.47
DreamDecoding Strategy=Low Confidence, Diffusion Steps=1282026.020.4685
Base + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=0%, Supervised Fine-Tuning (SFT)=true2026.060.451
LLaMA 1 33BActive Params=33B2024.010.441
Fisher-IntDim-G + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=50%, Supervised Fine-Tuning (SFT)=true2026.060.438
TransformerModel Scale=4B2026.060.434
YOCO (Dense)Model Scale=4B2026.060.43
DreamDecoding Strategy=Low Confidence, Diffusion Steps=642026.020.4011
DreamDecoding Strategy=Random, Diffusion Steps=1282026.020.395
Expert-level Fisher + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=50%, Supervised Fine-Tuning (SFT)=true2026.060.391
DreamDecoding Strategy=Random, Diffusion Steps=2562026.020.3836
DreamDecoding Strategy=Low Confidence, Diffusion Steps=2562026.020.3836
MoE comp. (Fisher) + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=50%, Supervised Fine-Tuning (SFT)=true2026.060.371
DreamDecoding Strategy=Random, Diffusion Steps=642026.020.3601
DreamDecoding Strategy=Margin, Diffusion Steps=642026.020.357
LLaMA 2 13BActive Params=13B2024.010.343
DreamDecoding Strategy=Entropy, Diffusion Steps=642026.020.3397
MoE comp. (activation) + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=50%, Supervised Fine-Tuning (SFT)=true2026.060.318
MoE comp. + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=50%, Supervised Fine-Tuning (SFT)=true2026.060.306
SFTModel backbone=Qwen3-8B, Fine-tuning method=SFT2026.050.293
SFTModel backbone=Qwen3-4B-It, Fine-tuning method=SFT2026.050.234
LLaMA 2 7BActive Params=7B2024.010.16
MoE-Pruner + SFTBackbone Model=OLMoE-1B-7B-0125, Compression Ratio (p)=50%, Supervised Fine-Tuning (SFT)=true2026.060.152
SFTModel backbone=Qwen3-1.7B, Fine-tuning method=SFT2026.050.098
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=02026.050.058
OPSDModel backbone=Qwen3-1.7B, Fine-tuning method=OPSD2026.050.034
SFTBackbone=Qwen3-4B2026.020.0053