Could On-Policy Self-Distillation Revolutionize Post-Training?
中文摘要
基于策略的自我蒸馏技术为大模型后训练提供新方案,有望在不破坏原有能力的前提下高效赋予模型新技能。
English Summary
On-policy self-distillation offers a novel reinforcement learning approach to effectively teach foundation models new skills without compromising their existing capabilities.
Original Excerpt
Teaching foundation models new skills without breaking them is hard. A new reinforcement learning-based approach might have the power to change that.