Classification on CB
91.1AccuracyMeZO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MeZOModel Backbone=Llama-3-8B, Model Precision=16 bits, Memory Profiling=20.5GB2025.05 | 91.1 | — | |
| AR1-ZOModel=Qwen3-32B, Rank (r)=642026.05 | 90.4 | — | |
| MeZO-LoRAModel=Qwen3-32B, Rank (r)=642026.05 | 89.3 | — | |
| AR1-ZOModel=Qwen3-1.7B, Rank (r)=642026.05 | 87.5 | — | |
| First Order Adamw FTModel Backbone=Llama2-7B, Fine-tuning Strategy=Full Fine-tuning, Optimization Algorithm=First Order Adamw, Number of training examples=10002026.06 | 85.71 | — | |
| SubZero-GV (Prefix)Optimization Algorithm=SubZero, Tuning Strategy=Prefix Tuning, Guiding Vectors=true, Model=OPT-13B2026.01 | 85.7 | — | |
| FT (Adam)Model=OPT-2.7B2026.05 | 85.7 | — | |
| FT (Adam)Model=OPT-13B2026.05 | 84.1 | — | |
| FTOptimization Algorithm=Adam, Tuning Strategy=Full Tuning, Model=OPT-13B2026.01 | 83.9 | — | |
| SubZero (Prefix)Optimization Algorithm=SubZero, Tuning Strategy=Prefix Tuning, Guiding Vectors=false, Model=OPT-13B2026.01 | 80.3 | — | |
| SubZero-GV (FT)Optimization Algorithm=SubZero, Tuning Strategy=Full Tuning, Guiding Vectors=true, Model=OPT-13B2026.01 | 73.9 | — | |
| SubZero-GV (LoRA)Optimization Algorithm=SubZero, Tuning Strategy=LoRA, Guiding Vectors=true, Model=OPT-13B2026.01 | 73.8 | — | |
| SubZero (FT)Optimization Algorithm=SubZero, Tuning Strategy=Full Tuning, Guiding Vectors=false, Model=OPT-13B2026.01 | 73.2 | — | |
| MeZO-GV (Prefix)Optimization Algorithm=MeZO, Tuning Strategy=Prefix Tuning, Guiding Vectors=true, Model=OPT-13B2026.01 | 73.2 | — | |
| Fine-tuningModel Backbone=OPT-6.7B, Model Precision=16 bits, Memory Profiling=26.8GB2025.05 | 73.2 | — | |
| MeZO-LoRAModel=Qwen3-1.7B, Rank (r)=642026.05 | 73.2 | — | |
| MeZO-GV (LoRA)Optimization Algorithm=MeZO, Tuning Strategy=LoRA, Guiding Vectors=true, Model=OPT-13B2026.01 | 72.8 | — | |
| ZO-AdaMU (Prefix)Optimization Algorithm=ZO-AdaMU, Tuning Strategy=Prefix Tuning, Model=OPT-13B2026.01 | 72.3 | — | |
| ZO-AdaMU (LoRA)Optimization Algorithm=ZO-AdaMU, Tuning Strategy=LoRA, Model=OPT-13B2026.01 | 71.6 | — | |
| MeZO-GV (FT)Optimization Algorithm=MeZO, Tuning Strategy=Full Tuning, Guiding Vectors=true, Model=OPT-13B2026.01 | 71.6 | — | |
| MeZO FTModel Backbone=Llama2-7B, Fine-tuning Strategy=Full Fine-tuning, Optimization Algorithm=MeZO, Number of training examples=10002026.06 | 71.43 | — | |
| MeZO(FT)Backbone=OPT-1.3B, Optimization Method=MeZO, Adaptation Technique=Fine-tuning2026.01 | 71.4 | — | |
| SubZero (LoRA)Optimization Algorithm=SubZero, Tuning Strategy=LoRA, Guiding Vectors=false, Model=OPT-13B2026.01 | 71.4 | — | |
| MeZO(LoRA)Backbone=OPT-1.3B, Optimization Method=MeZO, Adaptation Technique=LoRA2026.01 | 71.2 | — | |
| MeZO-GV(Prefix)Backbone=OPT-1.3B, Optimization Method=MeZO, Adaptation Technique=Prefix Tuning, Guiding Vectors (GV)=true2026.01 | 70.9 | — | |
| MeZO-GV(LoRA)Backbone=OPT-1.3B, Optimization Method=MeZO, Adaptation Technique=LoRA, Guiding Vectors (GV)=true2026.01 | 70.5 | — | |
| MeZO LoRAModel Backbone=Llama2-7B, Fine-tuning Strategy=LoRA, Optimization Algorithm=MeZO, Number of training examples=10002026.06 | 69.64 | — | |
| Dominant-layer ZO FTModel Backbone=Llama2-7B, Fine-tuning Strategy=Full Fine-tuning, Optimization Algorithm=Dominant-layer ZO, Number of training examples=10002026.06 | 69.64 | — | |
| MeZO-GV(FT)Backbone=OPT-1.3B, Optimization Method=MeZO, Adaptation Technique=Fine-tuning, Guiding Vectors (GV)=true2026.01 | 69.6 | — | |
| MeZO(Prefix)Backbone=OPT-1.3B, Optimization Method=MeZO, Adaptation Technique=Prefix Tuning2026.01 | 69.6 | — | |
| HiZOO (LoRA)Optimization Algorithm=HiZOO, Tuning Strategy=LoRA, Model=OPT-13B2026.01 | 69.6 | — | |
| HiZOO (Prefix)Optimization Algorithm=HiZOO, Tuning Strategy=Prefix Tuning, Model=OPT-13B2026.01 | 69.6 | — | |
| MeZO (Prefix)Optimization Algorithm=MeZO, Tuning Strategy=Prefix Tuning, Guiding Vectors=false, Model=OPT-13B2026.01 | 69.6 | — | |
| QZOModel Backbone=Llama-2-7B, Model Precision=4 bits, Memory Profiling=5.0GB2025.05 | 69.6 | — | |
| QZOModel Backbone=Llama-3-8B, Model Precision=4 bits, Memory Profiling=6.3GB2025.05 | 69.6 | — | |
| AR1-ZOModel=OPT-2.7B, Rank (r)=642026.05 | 69.6 | — | |
| LOZOModel=OPT-13B2026.05 | 69.6 | — | |
| AR1-ZOModel=OPT-13B, Rank (r)=642026.05 | 69.6 | — | |
| MeZOBackbone=Qwen-2.5-1.5B, Precision=BF162026.05 | 69.5 | — | |
| HiZOOOptimization Algorithm=HiZOO, Tuning Strategy=Full Tuning, Model=OPT-13B2026.01 | 69.4 | — | |
| ZO-AdaMU (2x)Optimization Algorithm=ZO-AdaMU, Tuning Strategy=Full Tuning, Training Steps=2x, Model=OPT-13B2026.01 | 67.9 | — | |
| MeZO (FT)Optimization Algorithm=MeZO, Tuning Strategy=Full Tuning, Guiding Vectors=false, Model=OPT-13B2026.01 | 67.9 | — | |
| MeZOModel Backbone=OPT-6.7B, Model Precision=16 bits, Memory Profiling=14.8GB2025.05 | 67.9 | — | |
| QZOModel Backbone=OPT-6.7B, Model Precision=4 bits, Memory Profiling=4.8GB2025.05 | 67.9 | — | |
| MeZOModel Backbone=Llama-2-7B, Model Precision=16 bits, Memory Profiling=14.8GB2025.05 | 67.9 | — | |
| MeZOBackbone=Llama-2-7B, Precision=BF162026.05 | 67.9 | — | |
| LOZOModel=OPT-2.7B2026.05 | 67.9 | — | |
| Dominant-layer ZO LoRAModel Backbone=Llama2-7B, Fine-tuning Strategy=LoRA, Optimization Algorithm=Dominant-layer ZO, Number of training examples=10002026.06 | 67.86 | — | |
| MeZO (LoRA)Optimization Algorithm=MeZO, Tuning Strategy=LoRA, Guiding Vectors=false, Model=OPT-13B2026.01 | 66.1 | — | |
| MeZO-LoRAModel=OPT-2.7B, Rank (r)=642026.05 | 66.1 | — | |
| MeZO-LoRAModel=OPT-13B, Rank (r)=642026.05 | 66.1 | — | |
| Fine-tuningModel Backbone=Llama-3-8B, Model Precision=16 bits, Memory Profiling=31.9GB2025.05 | 62.5 | — | |
| Fine-tuningModel Backbone=Llama-2-7B, Model Precision=16 bits, Memory Profiling=26.0GB2025.05 | 60.7 | — | |
| CAQ-ZOBackbone=Qwen-2.5-1.5B, Precision=NF42026.05 | 60.5 | — | |
| CAQ-ZOBackbone=Llama-2-7B, Precision=NF42026.05 | 58.7 | — | |
| ICLEvaluation Protocol=In-context Learning, Model=OPT-13B2026.01 | 57.1 | — | |
| QZOModel=Llama-2-13B, Precision=2 bits, Memory Profiling=5.78GB, Evaluation Protocol=Zeroth-order optimization2025.05 | 55.4 | — | |
| Zero-ShotModel Backbone=OPT-6.7B, Model Precision=16 bits2025.05 | 51.8 | — | |
| Zero-Shot-QModel Backbone=OPT-6.7B, Model Precision=4 bits2025.05 | 51.8 | — | |
| Zero-shotEvaluation Protocol=Zero-shot, Model=OPT-13B2026.01 | 46.4 | — | |
| Zero-Shot-QModel=Llama-2-13B, Precision=2 bits, Memory Profiling=N/A, Evaluation Protocol=Zero-shot2025.05 | 46.4 | — | |
| Zero-ShotModel Backbone=Llama-3-8B, Model Precision=16 bits2025.05 | 46.4 | — | |
| ZO-Alt-NaiveModel=Qwen3-32B2026.05 | 45.4 | — | |
| ICLBackbone=OPT-1.3B, Learning Strategy=In-context Learning2026.01 | 44.6 | — | |
| QuZOBackbone=Qwen-2.5-1.5B, Precision=NF42026.05 | 41.7 | — | |
| ZO-Alt-NaiveModel=Qwen3-1.7B2026.05 | 41.1 | — | |
| Zero-shotBackbone=OPT-1.3B, Learning Strategy=Zero-shot2026.01 | 39.3 | — | |
| Zero-Shot-QModel Backbone=Llama-3-8B, Model Precision=4 bits2025.05 | 37.5 | — | |
| QuZOBackbone=Llama-2-7B, Precision=NF42026.05 | 37.3 | — | |
| Zero-Shot-QBackbone=Llama-2-7B, Precision=NF42026.05 | 36.4 | — | |
| Zero-Shot-QModel Backbone=Llama-2-7B, Model Precision=4 bits2025.05 | 35.7 | — | |
| ZO-Alt-NaiveModel=OPT-13B2026.05 | 33.7 | — | |
| Zero-shot w/o finetuneModel Backbone=Llama2-7B, Fine-tuning Strategy=None, Optimization Algorithm=None, Number of training examples=10002026.06 | 32.14 | — | |
| Zero-ShotModel Backbone=Llama-2-7B, Model Precision=16 bits2025.05 | 32.1 | — | |
| ZO-Alt-NaiveModel=OPT-2.7B2026.05 | 28.6 | — | |
| Zero-Shot-QBackbone=Qwen-2.5-1.5B, Precision=NF42026.05 | 17.8 | — |