DEEPSEEK R1 PAPER

Summary from the TLDR AI newsletter

DeepSeek has released a paper along with its strong MIT licensed reasoning models. Importantly, the paper shows that there is no need for extremely fancy process reward models. DeepSeek simply used accuracy of answer, a formatting reward, and large scale RL to encourage reasoning in the model. It found that distilling this reasoning ability via SFT to smaller models gave an incredible boost in final performance.