A LinkedIn post from Sakana AI highlights new research on test-time scaling techniques for diffusion language models, presented in its ICML 2026 paper “UnMaskFork.” The post explains that the UnMaskFork (UMF) method enables multiple masked diffusion language models to collaborate on a single answer, improving performance on coding and math tasks without additional training.
According to the post, masked diffusion language models generate text by gradually filling in a fully masked sequence, allowing parallel generation and full-sequence visibility. The company’s research suggests that traditional large language model strategies, such as increasing temperature for diversity, are less effective for these models, and instead proposes diversity through “model switching” supported by Monte Carlo Tree Search.
The LinkedIn content indicates that UMF leverages pre-trained models trained on different data and methods, combining them at inference time to boost performance and scalability. This approach aligns with Sakana AI’s broader focus on “collective intelligence of AI,” including related methods like AB-MCTS and Sakana Fugu, all of which explore multi-model collaboration.
For investors, the post suggests ongoing advancement in inference-time optimization and ensemble techniques that could enhance the competitiveness of Sakana AI’s model offerings, particularly for complex coding and math applications. If these methods translate into superior performance or efficiency in commercial products, they may strengthen the company’s positioning in the high-performance AI segment and support future monetization opportunities in enterprise and developer markets.