OPDArena maps learning, forgetting in OPD

OPDArena’s strongest contribution is treating forgetting as a first-class post-training metric, not an afterthought to benchmark gains. The results suggest teacher provenance and student–teacher compatibility may matter as much as the distillation objective itself.

• –Evaluates five training methods across five math benchmarks and seven general-capability benchmarks
• –Finds RLVR-trained teachers produce more targeted student updates, while SFT-trained teachers cause greater capability loss
• –Shows out-of-domain distribution changes correlate more consistently with forgetting than in-domain changes
• –Includes reproducible training, evaluation, JS/KL analysis, parameter geometry, and teacher-source controls
• –Raises the practical bar: post-training pipelines should report retention and distribution shift alongside target-task accuracy