Scalable Oversight on SWE-bench Lite
6.27Misalignment RateCCO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CCOTarget α=0.05, Selective misalignment probability (p)=0.3, Candidate utilities=fixed, Number of seeds=52026.05 | 6.27 | 1.27 | 78.7 | 66.47 | |
| CCOTarget α=0.10, Selective misalignment probability (p)=0.3, Candidate utilities=fixed, Number of seeds=52026.05 | 10.87 | 0.87 | 63.28 | 46.67 | |
| CCOTarget α=0.15, Selective misalignment probability (p)=0.3, Candidate utilities=fixed, Number of seeds=52026.05 | 15.67 | 0.67 | 46.29 | 34.27 | |
| CCOTarget α=0.20, Selective misalignment probability (p)=0.3, Candidate utilities=fixed, Number of seeds=52026.05 | 19.53 | 0.47 | 32.76 | 21.13 |