LC-GRPO adds a Langevin correction step after each ODE rollout step so stochastic RL training rollouts stay close to deterministic test-time samples, improving flow-model reward optimization.
gummy bear zone
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LC-GRPO: Bridging Train-Inference Gap for Flow-Based GRPO with Langevin Correction
LC-GRPO adds a Langevin correction step after each ODE rollout step so stochastic RL training rollouts stay close to deterministic test-time samples, improving flow-model reward optimization.