How to Guide Your Language Flow

Apple Machine Learning Research introduced a method called probe guidance for guiding flow matching models. According to the published description, the approach uses the frozen internal states of an existing diffusion model to construct a guidance signal.

The method works on a principle similar to autoguidance, but it eliminates the need for an additional forward pass at inference time and provides a path to ensure that weak and strong models share similar dynamics, the authors state.

The authors are Rohit Dilip, Tianrong Chen, Yuyang Wang, David Van Valen, Josh Susskind, and Miguel Angel Bautista. Dilip and Van Valen are affiliated with Caltech, and the work was done while Dilip was at Apple.

Probe guidance was applied and benchmarked on continuous diffusion language models, where it sets a new state-of-the-art performance on unconditional generation, according to the research summary. When applied to a 1.7B diffusion language model, probe guidance consistently improves results on multiple choice question answering benchmarks.

Using the probes, the authors also studied the traditional autoguidance setting in which the strong model is a weak checkpoint. They report that the weak model must come from a low-entropy region of training. The findings are described as offering a practical way to improve diffusion language models and as shedding light on the mechanism behind autoguidance, which the authors say is currently poorly understood.