Software Engineering on SWE-bench Lite (300 instances)
0Misalignment RateAlways Baseline
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Always Baselinealpha (α)=0, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 0 | 12 | |
| CCOalpha (α)=0.01, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 1.3 | 13 | |
| CCOalpha (α)=0.05, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 5 | 14.3 | |
| CCOalpha (α)=0.10, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 9.7 | 15.3 | |
| CCOalpha (α)=0.15, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 14 | 16.7 | |
| CCOalpha (α)=0.20, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 17 | 17.7 | |
| CCOalpha (α)=0.25, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 21 | 19.7 | |
| Unconstrainedlambda (λ)=0, Agent=Claude Sonnet 4.6, Misalignment probability (p)=0.5, Step size (η)=0.32026.05 | 55 | 29 |