specific method · filed under post-training
Direct RL optimization of a generative reward model
Applies reinforcement-learning optimization directly to the generative reward model.
- source
- 1
- models
- 2
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
Crucially, we apply RL optimization directly to the GRM itself.
usedpost trainingin DeepSeek-V4DeepSeek
Filed alongside
Other methods under post-training :: reward modelling.
Generative Reward ModelGroupwise Agentic GradingGroupwise Reward SynthesisAdversarial screeningLength-adjusted RL rewardVerifier cross-checkingAbstention-aware reward for factual QAAgentic Generative Reward ModelBehavior rubricsBinary task verifierBinary terminal-verifier rewardCollaboration bonusDeterministic chain of checkersFive-dimension comparative grading of passing patchesHack-agent screeningHybrid reward systemLanguage consistency rewardMonitoring-only penalty strategyMulti-level reward formulationMultiplicative reward synthesisNegative checks for unintended side effectsOutcome Reward ModelPer-token tool-error reward shapingPrinciple-conditioned Generative Reward Model