OPEN LINK ↗
// 1d agoRESEARCH PAPER
UnMaskFork scales test-time compute for MDLMs
UnMaskFork is a new test-time scaling framework that formulates the unmasking trajectory of Masked Diffusion Language Models (MDLMs) as a search tree. By utilizing Monte Carlo Tree Search with deterministic partial unmasking actions, the framework achieves efficient state space exploration and outperforms scaling baselines on coding and math reasoning benchmarks.
// ANALYSIS
Applying search to diffusion-based generation has traditionally been held back by stochastic path generation, but UnMaskFork's deterministic branching elegantly adapts MCTS to this architecture.
- –**Deterministic Action Branching** eliminates the random noise transitions of typical diffusion generation, allowing structured search algorithms to systematically evaluate alternative generation paths.
- –**State Caching Efficiency** is greatly enhanced since deterministic transitions enable caching of early unmasking states across different search branches.
- –**Competitive Reasoning Performance** on coding and math tasks suggests that diffusion-based models, when scaled at test time, can offer viable alternatives to standard autoregressive reasoning models.
// TAGS
masked-diffusion-language-modelstest-time-scalingmonte-carlo-tree-searchmachine-learninggenerative-ai
DISCOVERED
1d ago
2026-07-21
PUBLISHED
1d ago
2026-07-21
RELEVANCE
7/ 10
AUTHOR
hardmaru