T$^\star$: Progressive Block Scaling for MDM Through Trajectory Aware RL

Abstract: We present T$^\star$, a simple \textsc{TraceRL}-based training curriculum for progressive block-size scaling in masked diffusion LLMs (MDMs). Starting from an AR-initialized small-block MDM, T$^\star$~transitions smoothly to larger blocks, enabling higher-parallelism decoding with minimal performance degradation on math reasoning benchmarks. Moreover, further analysis suggests that T$^\star$~can converge to an alternative decoding schedule $\hat{\rm S}$ that achieves comparable performance.