method · first seen 5 Nov 2026
group-relative baseline subtraction
Also: group-mean baseline
Variance-reducing baseline in policy gradients computed as the mean reward over a group of sampled responses.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement LearningPapers · 5 Nov 2026 · Chenruo Liu, Zihan Wang, Yijun Dong, Qi Lei
Latest sources
- Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement LearningPapers · 5 Nov 2026 · Chenruo Liu, Zihan Wang, Yijun Dong, Qi Lei