Score: 0

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

Published: January 16, 2026 | arXiv ID: 2601.11214v1

By: Hanchen Xia , Baoyou Chen , Yutang Ge and more

Potential Business Impact:

Makes AI better at solving math problems faster.

Business Areas:
A/B Testing Data and Analytics

We present T$^\star$, a simple \textsc{TraceRL}-based training curriculum for progressive block-size scaling in masked diffusion language models (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.

Page Count
8 pages

Category
Computer Science:
Computation and Language