Knowledge Retention on MATH (retain)
37AccuracyRLVR source
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| RLVR source2026.06 | 37 | — | |
| MASTLearning rate (η)=8×10−6, Configuration=selected2026.06 | 28.5 | -8.5 | |
| Full-parameter GALearning rate (η)=5×10−62026.06 | 23 | -14 | |
| Full-parameter GALearning rate (η)=8×10−62026.06 | 9.5 | -27.5 |