Search papers, labs, and topics across Lattice.
1
0
2
Overtraining in SFT can lead to rank inversion, with GRPO performance plummeting despite improved pre-RL metrics, highlighting a critical failure mode in RLHF training.
