Research
-
02Technical ReportNoise Your Prompt: Improved Conditional Generation in Diffusion Language Models
We offer a simple modification to training continuous diffusion language models that leads to improvements in both accuracy and generation diversity on combinatorial reasoning benchmarks.
-
03Technical ReportLooped Flow Matching for Few‑Step Language Models
We improve continuous diffusion language models in the few-step regime by allowing them to correct previous integration errors, borrowing intuition from recursive reasoning models. On Sudoku hard this increases exact match accuracy from 38% to 72%.