Pairwise Judge Comparison on UltraFeedback core250
161Win Count (W)TEA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TEAm=32, selection=bo512, baseline=GRPO2026.05 | 161 | 64.4 | |
| TEAm=16, selection=bo128, baseline=GRPO2026.05 | 159 | 63.6 | |
| TEAselection=bo128, baseline=GRPO2026.05 | 152 | 60.8 | |
| TEAm=64, selection=bo128, baseline=GRPO2026.05 | 152 | 60.8 | |
| TEAm=32, selection=bo256, baseline=GRPO2026.05 | 148 | 59.2 | |
| TEAm=16, selection=bo512, baseline=GRPO2026.05 | 146 | 58.4 | |
| TEAselection=bo128, baseline=Best TTA2026.05 | 145 | 58 | |
| TEAm=64, selection=bo512, baseline=GRPO2026.05 | 145 | 58 | |
| TEAbackbone=Qwen3-4B, selection=bo128, baseline=GRPO2026.05 | 144 | 57.6 | |
| TEAm=16, selection=bo256, baseline=GRPO2026.05 | 141 | 56.4 | |
| TEAm=32, selection=bo128, baseline=GRPO2026.05 | 141 | 56.4 | |
| TEAm=64, selection=bo256, baseline=GRPO2026.05 | 141 | 56.4 | |
| TEAbackbone=Llama-3.1-8B LoRA, selection=bo128, baseline=GRPO2026.05 | 137 | 54.8 | |
| TEAselection=bo128, baseline=GRPO-Z2026.05 | 136 | 54.4 |