T$^\star$: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learning
Title: T$^\star$: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learning
Abstract: This paper introduces T$^\star$, a straightforward TraceRL-based training curriculum designed for the progressive scaling of block sizes in masked diffusion language models (MDMs). Beginning with an MDM initialized using autoregressive (AR) methods and configured for small blocks, T$^\star$ facilitates a seamless transition to larger block sizes. This approach allows for decoding with significantly higher parallelism while maintaining minimal performance loss on mathematical reasoning benchmarks. Additionally, our further analysis indicates that T$^\star$ might converge toward an alternative decoding schedule that delivers comparable results.
Source: arXiv Generated at: 2026-06-04 00:00:00 UTC
