New Method Enables Policy Gradient Post-Training for Masked Diffusion Language Models
Researchers propose a variance-reduced post-training approach that makes policy gradient methods usable for masked diffusion language models, where required log-likelihoods are hard to compute.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-16 05:00 · arXiv cs.CL
See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion Language Models