Self-correction on Qwen-72B math failure pool n=30
80Correction RateL_MEMORY (Source-Conditioned Role Relabeling)
Evaluation Results
| Method | Links | |
|---|---|---|
| L_MEMORY (Source-Conditioned Role Relabeling)Protocol=L_MEM.2026.06 | 80 | |
| ReflexionProtocol=Reflex.2026.06 | 46.7 | |
| auditProtocol=audit-only baseline2026.06 | 30 | |
| Chain-of-Verification (CoVe)Protocol=CoVe2026.06 | 10 | |
| Self-RefineProtocol=S.-Refine2026.06 | 6.7 |