Creative Writing on WildBench
83.9WildBench ScoreBaseline
Evaluation Results
| Method | Links | |
|---|---|---|
| BaselineModel=GLM-4.5-Air, Compression=Baseline2025.10 | 83.9 | |
| REAPModel=GLM-4.5-Air, Compression=25%, Technique=Pruning2025.10 | 83.1 | |
| EANModel=GLM-4.5-Air, Compression=25%, Technique=Pruning2025.10 | 82.4 | |
| BaselineModel=Qwen3-30B-A3B, Compression=Baseline2025.10 | 81.1 | |
| EANModel=Qwen3-30B-A3B, Compression=25%, Technique=Pruning2025.10 | 81.1 | |
| FrequencyModel=Qwen3-30B-A3B, Compression=25%, Technique=Pruning2025.10 | 80.7 | |
| M-SMoEModel=Qwen3-30B-A3B, Compression=25%, Technique=Merging2025.10 | 80.5 | |
| REAPModel=Qwen3-30B-A3B, Compression=25%, Technique=Pruning2025.10 | 80.4 | |
| FrequencyModel=GLM-4.5-Air, Compression=25%, Technique=Pruning2025.10 | 79.3 | |
| HC-SMoEModel=GLM-4.5-Air, Compression=25%, Technique=Merging2025.10 | 78.8 | |
| M-SMoEModel=GLM-4.5-Air, Compression=25%, Technique=Merging2025.10 | 78.1 | |
| REAPModel=GLM-4.5-Air, Compression=50%, Technique=Pruning2025.10 | 75.4 | |
| M-SMoEModel=Qwen3-30B-A3B, Compression=50%, Technique=Merging2025.10 | 72.5 | |
| REAPModel=Qwen3-30B-A3B, Compression=50%, Technique=Pruning2025.10 | 71.8 | |
| EANModel=Qwen3-30B-A3B, Compression=50%, Technique=Pruning2025.10 | 70.2 | |
| EANModel=GLM-4.5-Air, Compression=50%, Technique=Pruning2025.10 | 70.2 | |
| FrequencyModel=Qwen3-30B-A3B, Compression=50%, Technique=Pruning2025.10 | 67.7 | |
| FrequencyModel=GLM-4.5-Air, Compression=50%, Technique=Pruning2025.10 | 60.4 | |
| HC-SMoEModel=GLM-4.5-Air, Compression=50%, Technique=Merging2025.10 | 59.3 | |
| HC-SMoEModel=Qwen3-30B-A3B, Compression=25%, Technique=Merging2025.10 | 49.7 | |
| FullModel=ERNIE-21B, Pruning Ratio=0%, Protocol=Zero-shot2026.03 | 47.9 | |
| SCOPEBase Model=Qwen3-8B, Task Type=Open-ended, # Seed Prompts=0, Rubric Source=Self-generated, Judge=Self-judge, Training Recipe=GRPO (60 steps), External Model Dependency=None2026.05 | 46.1 | |
| FullModel=OLMoE-7B, Pruning Ratio=0%, Protocol=Zero-shot2026.03 | 44.4 | |
| WebExplorerBase Model=Qwen3-8B, Task Type=Short-form QA, # Seed Prompts=~12K, Rubric Source=—, Judge=—, Training Recipe=SFT → GRPO (∼380 steps), External Model Dependency=DeepSeek-V3, Gemini 2.5 Flash2026.05 | 41.7 | |
| REAPModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 40.5 | |
| AIMERModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 40.2 | |
| EANModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 39.6 | |
| M-SMoEModel=GLM-4.5-Air, Compression=50%, Technique=Merging2025.10 | 39.1 | |
| SEERModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 38.1 | |
| RandomModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 37.7 | |
| WebThinker-R1Base Model=QwQ-32B, Task Type=Short-form QA, # Seed Prompts=~3K, Rubric Source=—, Judge=—, Training Recipe=DPO (2 iters), External Model Dependency=Qwen2.5-72B-Instruct2026.05 | 36.5 | |
| DR TuluBase Model=Qwen3-8B, Task Type=Deep research, # Seed Prompts=~9K, Rubric Source=GPT-4.1, Judge=GPT-4.1-mini, Training Recipe=SFT → GRPO (1.9K steps), External Model Dependency=GPT-4.1, GPT-4.1-mini, GPT-52026.05 | 36 | |
| FrequencyModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 35.2 | |
| RandomModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 29.7 | |
| REAPModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 26 | |
| EANModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 25.3 | |
| AIMERModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 23.6 | |
| MagnitudeModel=ERNIE-21B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 22.8 | |
| AIMERModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 21.7 | |
| EANModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 20.5 | |
| REAPModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 19.9 | |
| SEERModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 17.4 | |
| FrequencyModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 16.4 | |
| RandomModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 15.8 | |
| SEERModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 14.1 | |
| FrequencyModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 12.7 | |
| MagnitudeModel=ERNIE-21B, Pruning Ratio=50%, Protocol=Zero-shot2026.03 | 7.7 | |
| MagnitudeModel=OLMoE-7B, Pruning Ratio=25%, Protocol=Zero-shot2026.03 | 2 | |
| HC-SMoEModel=Qwen3-30B-A3B, Compression=50%, Technique=Merging2025.10 | 0.8 |