Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT? 文章

ArXiv CS.CL2026-05-29NEWSen作者: Jeanmely Rojas Nunez, Viraj Sawant, Nathan Allen, Nomgondalai Amgalanbaatar, Yannis Zongo, Vasu Sharma, Maheep Chaudhary

查看原文 →

Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT? · 相关技术

相关技术

policy-gradient large language models differential circuit vulnerability Policy Gradient Supervised Fine-Tuning reinforcement learning