Format Online paper / PDF
Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER How Code Diverges From Prose

"DeepSeek-R1 combines rule-based rewards for verifiable reasoning tasks with model-based rewards for more general tasks."

The researchers distinguish accuracy and formatting rewards for mathematics, coding, and logical reasoning from model-based evaluation of more general responses. Their reward architecture demonstrates why objectively testable problems provide a different optimization signal from open-ended language-generation tasks.

View Source
Source title DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
Source authors Daya Guo et al.
Source year 2025
Confidence high
Importance high
Back to evidence PREVIOUSNEXT