Salesforce: On-Policy Distillation Beats GRPO in Compute Efficiency

A Salesforce paper finds on-policy distillation from a self-teacher can save about 50x FLOPs versus GRPO at equal compute, and merging domain experts cuts another 42% of training cost.

2026-11-08 ~ 2026-11-08 · 2 related posts

• Salesforce paper: RL-trained experts merged via OPD save 42% FLOPs vs external-teacher distillation — rohanpaul_ai · 2026-11-08
• arXiv paper unpacks OPD vs GRPO: self-teacher distillation cuts FLOPs 50x, expert merging saves 42% — rohanpaul_ai · 2026-11-08