LLM-as-a-judge on HH-RLHF
81.3CoverageHeuristic Selection
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Heuristic SelectionCascade=M → L → O, Target agreement level=0.852026.05 | 81.3 | 0 | |
| Heuristic SelectionCascade=L → Q → O, Target agreement level=0.852026.05 | 72.4 | 0 | |
| CSE + PredictiveCascade=M → L → O, Target agreement level=0.852026.05 | 72.3 | 0 | |
| CSE + PredictiveCascade=L → Q → O, Target agreement level=0.852026.05 | 58.6 | 0 | |
| CSE + OursCascade=M → L → O, Target agreement level=0.852026.05 | 47.2 | 42.8 | |
| CSE + RandomCascade=M → L → O, Target agreement level=0.852026.05 | 45.4 | 29.8 | |
| CSE + SimulatedCascade=M → L → O, Target agreement level=0.852026.05 | 42.5 | 37 | |
| CSE + VanillaCascade=M → L → O, Target agreement level=0.852026.05 | 41.3 | 33.4 | |
| CSE + OursCascade=L → Q → O, Target agreement level=0.852026.05 | 36.9 | 51.6 | |
| CSE + RandomCascade=L → Q → O, Target agreement level=0.852026.05 | 34.2 | 35.7 | |
| CSE + VanillaCascade=L → Q → O, Target agreement level=0.852026.05 | 32.7 | 37.9 | |
| CSE + SimulatedCascade=L → Q → O, Target agreement level=0.852026.05 | 31 | 46.9 |