Instruction Following on AlpacaEval
98.4Win RateNone
Evaluation results
| Method | Links | ||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| NoneModel=Gemini-3-flash2026.05 | 98.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SelfdefenseModel=Claude-4.5-haiku2026.05 | 97.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoneModel=GPT-5-chat2026.05 | 97.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EvoDefenseModel=Gemini-3-flash2026.05 | 97.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BFPOBackbone=Llama3-8B-Instruct2025.12 | 97.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoneModel=Claude-4.5-haiku2026.05 | 96.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SelfdefenseModel=GPT-5-chat2026.05 | 96.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IBProtectorModel=GPT-5-chat2026.05 | 96.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Llama3-8B-Instruct2025.12 | 96.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCANBackbone=Llama3-8B-Instruct2025.12 | 95.71 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TeacherModel=Qwen2.5-14B-Instruct2025.10 | 95.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCANBackbone=Qwen2.5-7B-Instruct2025.12 | 95.28 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BFPOBackbone=Qwen2.5-7B-Instruct2025.12 | 95.22 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EvoDefenseModel=GPT-5-chat2026.05 | 95.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SelfdefenseModel=Gemini-3-flash2026.05 | 95.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPO-SBackbone=Qwen2.5-7B-Instruct2025.12 | 94.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPO-HBackbone=Qwen2.5-7B-Instruct2025.12 | 94.41 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IBProtectorModel=Gemini-3-flash2026.05 | 94.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-7B-Instruct2025.12 | 94.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PeCANBackbone=Qwen2.5-7B-Instruct2025.12 | 94.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NSPOBackbone=Qwen2.5-7B-Instruct2025.12 | 94.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPO-MixBackbone=Llama3-8B-Instruct2025.12 | 93.85 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TeacherModel=Qwen2.5-7B-Instruct, Role=Teacher2025.10 | 93.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPO-SBackbone=Llama3-8B-Instruct2025.12 | 93.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPO-MixBackbone=Qwen2.5-7B-Instruct2025.12 | 93.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EvoDefenseModel=Claude-4.5-haiku2026.05 | 93.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NSPOBackbone=Llama3-8B-Instruct2025.12 | 93.23 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-7BBackbone=Qwen2.5-7B2025.06 | 93.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Semantic SmoothModel=Claude-4.5-haiku2026.05 | 92.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Semantic SmoothModel=Gemini-3-flash2026.05 | 92.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoneModel=Mistral-7B-v0.22026.05 | 92.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Self DefendModel=Claude-4.5-haiku2026.05 | 91.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPO-HBackbone=Llama3-8B-Instruct2025.12 | 91.61 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| W-DOORBackbone=Llama3-8B-Instruct2025.12 | 91.61 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MT (Teacher)Distillation Pair (Teacher Model -> Student Model)=Qwen2.5-7B-Inst -> Qwen2.5-1.5B2026.04 | 91.43 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EvoDefenseModel=Mistral-7B-v0.22026.05 | 91.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SmoothLLMModel=Gemini-3-flash2026.05 | 90.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Self DefendModel=GPT-5-chat2026.05 | 89.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IBProtectorModel=Claude-4.5-haiku2026.05 | 88.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Self DefendModel=Gemini-3-flash2026.05 | 88.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MT (Teacher)Distillation Pair (Teacher Model -> Student Model)=Qwen2-7B-Inst -> Qwen2-1.5B2026.04 | 88.54 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PeCANBackbone=Llama3-8B-Instruct2025.12 | 88.45 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Llama-2-ChatParameter Count=7B2024.02 | 87.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Semantic SmoothModel=GPT-5-chat2026.05 | 87.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPOAlignment method=DPO, Base Model=LLaVA-1.5-13b2024.02 | 86.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SmoothLLMModel=GPT-5-chat2026.05 | 86.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SmoothLLMModel=Claude-4.5-haiku2026.05 | 85.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Self DefendModel=Mistral-7B-v0.22026.05 | 85.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TÜLU 2+DPOParameter Count=7B2024.02 | 85.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Aligned SFTModel=Llama2-13B, Setting=w/ training2023.11 | 83 | 33.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Goal Prioritization w/o thoughtsModel=Llama2-13B, Setting=w/ training2023.11 | 82 | 33.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Goal Prioritization w/o harmModel=Llama2-13B, Setting=w/ training2023.11 | 82 | 32.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| daDPOBackbone=Qwen2.5-1.5B2025.06 | 81.49 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Vicuna-1.5-13bModel=Vicuna-1.5-13b2024.02 | 81.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DistiLLM-2α=-5, Teacher Model=Qwen2.5-14B-Instruct, Student Model=Qwen2.5-1.5B-Instruct2025.10 | 81.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Vanilla SFTModel=Llama2-7B, Setting=w/ training2023.11 | 81 | 31.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Aligned SFTModel=Llama2-7B, Setting=w/ training2023.11 | 81 | 31.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LLAVA-RLHF-13bModel=LLAVA-RLHF-13b2024.02 | 81 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DistiLLM-2alpha=-5, Base Model=Qwen2.5-1.5B-Instruct2025.10 | 80.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoneModel=LLaMA3-8B-instruct2026.05 | 80.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DistiLLM-2α=-1, Teacher Model=Qwen2.5-14B-Instruct, Student Model=Qwen2.5-1.5B-Instruct2025.10 | 80.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Vanilla SFTModel=Llama2-13B, Setting=w/ training2023.11 | 80 | 32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NoneModel=Qwen-7B-chat2026.05 | 79.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| IBProtectorModel=LLaMA3-8B-instruct2026.05 | 79.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DistiLLM-2alpha=-1, Base Model=Qwen2.5-1.5B-Instruct2025.10 | 79.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RPOModel=Qwen2.5-7B, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 79.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Warmup-Stable-Only (WSO)Model=1B, αpre=1.0, αmid=1.02026.03 | 79.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WSDModel=1B, αpre=1.0, αmid=0.02026.03 | 79.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LLaVA-1.5-13bModel=LLaVA-1.5-13b2024.02 | 79.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Goal PrioritizationModel=Llama2-13B, Setting=w/ training2023.11 | 79 | 32.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| rDPOBackbone=Qwen2.5-1.5B2025.06 | 78.94 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RPOModel=Qwen2.5-7B, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 78.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| dDPOBackbone=Qwen2.5-1.5B2025.06 | 78.85 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RPOModel=LLaMA-3-8B, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 78.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WSDModel=1B, αpre=0.1, αmid=1.02026.03 | 78.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LLAMA-70B (QLORA)Model Training Status=Finetuned, Strategy=QLoRA2024.03 | 78.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GIFTModel Size=32B2025.10 | 78.53 | — | — | — | 61.43 | — | — | — | — | — | — | — | 3,327 | — | — | — | — | — | — | — | — | — | |
| IPOBackbone=Mistral-7B, Framework=DPO2025.02 | 78.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EvoDefenseModel=LLaMA3-8B-instruct2026.05 | 78.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RPOModel=LLaMA-3-8B, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 78.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NovelSelectBase Model=LLaMA-3-8B, Data Budget=10k samples2025.02 | 78.07 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Goal Prioritization w/o thoughtsModel=Llama2-7B, Setting=w/ training2023.11 | 78 | 31.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EvoDefenseModel=Qwen-7B-chat2026.05 | 78 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RPOModel=Mistral-7B, Baseline=SFT, LLM Judge=GPT-4o2025.06 | 77.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Mistral-7B-RewardBackbone=Mistral-7B, Preference Acquisition=Reward model-based classification2025.02 | 77.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Goal PrioritizationModel=Llama2-7B, Setting=w/ training2023.11 | 77 | 31.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RPOModel=Mistral-7B, Baseline=SFT, LLM Judge=Claude-3.52025.06 | 76.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WSDModel=1B, αpre=0.1, αmid=0.02026.03 | 76.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LinearModel=1B, αpre=0.1, αmid=1.02026.03 | 76.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VCRDDistillation Pair (Teacher Model -> Student Model)=Qwen2.5-7B-Inst -> Qwen2.5-1.5B2026.04 | 76.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Goal Prioritization w/o harmModel=Llama2-7B, Setting=w/ training2023.11 | 76 | 31.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| W-DOORBackbone=Qwen2.5-7B-Instruct2025.12 | 75.96 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QDITBase Model=LLaMA-3-8B, Data Budget=10k samples2025.02 | 75.91 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RandomBase Model=LLaMA-3-8B, Data Budget=10k samples2025.02 | 75.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| K-meansBase Model=LLaMA-3-8B, Data Budget=10k samples2025.02 | 75.46 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| K-Center-GreedyBase Model=LLaMA-3-8B, Data Budget=10k samples2025.02 | 75.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DistilLLMDistillation Pair (Teacher Model -> Student Model)=Qwen2.5-7B-Inst -> Qwen2.5-1.5B2026.04 | 75.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Self DefendModel=LLaMA3-8B-instruct2026.05 | 75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Rejection-samplingAlignment method=Rejection-sampling, Base Model=LLaVA-1.5-13b2024.02 | 74.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| dSFTBackbone=Qwen2.5-1.5B2025.06 | 74.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |