Test-time scaling -- the idea that you can make a model smarter just by giving it more compute at inference -- has become one of the most important levers in modern AI. It powers reasoning models, Best-of-N sampling, and MCTS-based search. But all of that work has been built around autoregressive LLMs, the kind that generate text one token at a time, left to right. A fundamentally different class of model has been quietly gaining ground, and nobody had figured out how to give it the same scaling treatment. Until now.
Sakana AI's UnMaskFork (UMF), accepted at ICML 2026, is the first test-time scaling method designed specifically for Masked Diffusion Language Models (MDLMs). Instead of generating text sequentially, MDLMs start with a fully masked sequence and iteratively fill in tokens in parallel, seeing the entire context at once. The paper shows that having multiple MDLMs collaborate on generating a single answer -- each picking up where the others left off -- consistently outperforms every existing test-time scaling baseline on coding benchmarks.
Why the Standard Playbook Fails Here
To understand why this matters, you need to know why the usual tricks don't work. In autoregressive models, you get diversity for free: raise the temperature (a parameter that flattens the probability distribution over tokens, making the model more willing to pick less-likely words), generate N different answers, and pick the best one. This is Best-of-N, and it's a workhorse of test-time scaling.
MDLMs like Dream-Coder and LLaDA don't play by those rules. The Sakana team tested Best-of-N on Dream-Coder and found that raising the temperature or randomizing the unmasking order sharply degrades answer quality. Best-of-N built on either source of randomness failed to surpass a single deterministically decoded answer. In other words, the very thing that makes Best-of-N work for GPT-style models -- controllable randomness -- actively breaks MDLM generation. The model needs to commit to high-confidence tokens first; injecting noise at that stage corrupts the whole sequence.
The Fork in the Road
UnMaskFork's insight is elegant: instead of creating diversity through randomness, create it through model switching
Don't miss what's next in AI
Join 300,000+ engineers and researchers who get the signal, not the noise.
- Full access to in-depth AI research breakdowns
- Be the first to know what's trending before it hits mainstream
- Daily curated papers, repos, and industry moves