[1/n] 🧐@deepseek_ai #DeepSeekR1 has shown the power of RL without SFT. But what does RL learns differently than SFT? Our answer is: 📉SFT Memorizes, RL Generalizes.📈

@robertarail Nice works! We are still trying to formulate some theory stuff lol

a pumping elephant paper

RT @ScuderiaFerrari: Welcoming Zhou Guanyu back to the Ferrari family as he joins Antonio Giovinazzi as our official reserve driver! Zhou s…

RT @oran_ge: Google 这篇论文的结论非常清晰：《SFT 负责记忆, RL 负责泛化》 简单结论： 监督微调 (SFT) 就像给学生看大量的例题和答案。 学生通过模仿例题来学习。 强化学习 (RL) 就像让学生自己解题，答对了给奖励，答错了给惩罚。 学生通过…

RT @YugeTen: ✨New blog post✨: my attempt as a vision researcher at finally understanding RLHF -- a deep dive into PPO & DeepSeek's GRPO! N…

@simon_zhai 🥳🥳🥳

@rosinality Thanks for your interest! Verification do improve OOD performance. But to clarify, baseline of fig. 10 is initial checkpoint’s accuracy with {1,3,5,10} verification. Hence direct contribution of verification has been counteracted.

RT @simon_zhai: [1/n] @deepseek_ai R1 has shown the power of RL without SFT. But what does RL learns differently than SFT? We have a answer…

I posted twice since twitter finds our research so sexy that it labeled the tweet as sensitive content…😅

Thanks to great collaborators @simon_zhai @jihanyang13 @TongPetersb and advisors @sainingxie Dale Schuurmans @quocleix @svlevine @YiMaTweets. 🎇Happy New Year!📷🐍