Showing cs.LGShow all
2 papers · 1 filter
cs.LG2026
GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs
Kai Yang, Jingwei Xu, Wanyu Wang +4
On-policy rollout methods such as GRPO are central to post-training of large language models, yet they frequently suffer from training instabilities, cross-task capability degradat…
cs.LG2026
Hidden Failure Modes of Gradient Modification under Adam in Continual Learning, and Adaptive Decoupled Moment Routing as a Repair
Yuelin Hu, Zhenbo Yu, Zhengxue Cheng +2
Many continual-learning methods modify gradients upstream (e.g., projection, penalty rescaling, replay mixing) while treating Adam as a neutral backend. We show this composition ha…