ReMaRL: Learning to Unmask and Remask in Frozen Diffusion Language Models
A roughly 300,000-parameter policy improved math and code results while leaving the underlying diffusion model frozen.
ReMaRL learns which tokens to unmask, retain or remask, allowing earlier predictions to be revised as new context appears. The policy is trained with Group Relative Policy Optimization using final-answer rewards. With LLaDA-8B-Instruct, it scored highest among the compared policies within the tested budgets: 4–32 model calls for math and 8–16 for code.
ArXiv · AI/CL/LG's note
ReMaRL learns which tokens to unmask, retain or remask, allowing earlier predictions to be revised as new context appears. The policy is trained with Group Relative Policy Optimization using final-answer rewards. With LLaDA-8B-Instruct, it scored highest among the compared policies within the tested budgets: 4–32 model calls for math and 8–16 for code.
ArXiv · AI/CL/LG's note
score 5