Apple proposes probe guidance to improve diffusion language models without extra inference cost

Apple’s Machine Learning Research team describes a new guidance method for flow matching models called probe guidance. Unlike prior approaches such as autoguidance, which require an extra forward pass during inference, probe guidance constructs a guidance signal using the frozen internal states of an existing diffusion model. This reduces computational overhead while maintaining or improving generation quality.

The researchers apply probe guidance to continuous diffusion language models and report state-of-the-art performance on unconditional generation benchmarks. When tested on a 1.7B parameter diffusion language model, the method consistently improves results on multiple-choice question answering benchmarks.

The paper also explores the relationship between probe guidance and autoguidance, particularly in cases where the strong model is a weak checkpoint. The findings suggest that the weak model must originate from a low-entropy region of training for effective guidance, offering new insight into a mechanism that has been poorly understood.

The work contributes both a practical technique for improving diffusion language models and a conceptual advance in understanding guidance dynamics in these systems.