Li et al.
Adaptive replay preserves reasoning diversity in RL fine-tuning
Targeted Adaptive Replay selectively restores suppressed reasoning modes that standard RLVR training loses.
Academic
Wen-Zhi Li · Sainyam Galhotra · Claire Cardie
Cornell University
Research Digest··3 min read
The authors show that reinforcement learning with verifiable rewards (RLVR) reduces reasoning diversity due to finite-sample dynamics, not the objective itself.
§