TY  - RPRT
TI  - What Will Post-Training Fix? Per-Problem Gains Are Shared Across Independent RL Runs, and Existing Checkpoints Predict Them Better Than A Priori Signals
AU  - Xiaoxian Duan
PY  - 2026
UR  - https://arxiv.org/abs/2610.04978
ID  - 2610.04978
ER  -