Apple Machine Learning researchers have introduced a new method called probe guidance to guide flow matching models. This approach uses the frozen internal states of an existing diffusion model to construct a guidance signal. According to the researchers, this method operates on a principle similar to autoguiding but eliminates the need for an additional forward pass at inference time, while ensuring that weak and strong models share similar dynamics.
Apple Research Introduces Probe Guidance to Improve Diffusion Language Models
In benchmarks on continuous diffusion language models, probe guidance set a new state-of-the-art performance for unconditional generation. When applied to a 1.7B diffusion language model, the method consistently improved performance on multiple-choice question-answering benchmarks.
Additionally, the study investigated the mechanism behind traditional autoguiding settings. The researchers found that for autoguiding to be effective, the weak model must originate from a low-entropy region of training. These findings provide a practical way to enhance diffusion language models and offer new insights into the mechanisms of autoguiding.
Sources
- How to Guide Your Language Flow （Apple Machine Learning, 2026-09-23）
- arXiv