Worst-case Performance Gap Analysis on Summarization & Email Writing Combined
0.1586Max SubOptLateEnsemble
Evaluation Results
| Method | Links | |
|---|---|---|
| LateEnsembleUser LLM=Qwen-3 32B, Seeds=32026.01 | 0.1586 | |
| EarlyEnsembleUser LLM=Qwen-3 32B, Seeds=32026.01 | 0.1612 | |
| SFTUser LLM=Qwen-3 32B, Seeds=32026.01 | 0.5772 | |
| BaseUser LLM=Qwen-3 32B, Seeds=32026.01 | 0.7364 | |
| DPOUser LLM=Qwen-3 32B, Seeds=32026.01 | 0.8698 |