Bug-fixing on HumanEvalFix (held-out 25% eval)
5.2Utility Difference (Δπ)bayesian_greedy
Evaluation Results
| Method | Links | |
|---|---|---|
| bayesian_greedyGenerator=gpt-5-mini2026.06 | 5.2 | |
| bayesian_greedyGenerator=claude-sonnet-4.52026.06 | 4.2 | |
| bayesian_DPGenerator=gpt-5-mini2026.06 | 0.1 |