Probe guidance, introduced by researchers at Caltech and Apple, is a method for guiding flow matching models without the extra inference-time forward pass that autoguidance requires. The guidance signal is constructed from the frozen internal states of an existing diffusion model. Autoguidance typically pairs a weak model with a strong model and needs two passes during generation; probe guidance instead derives the signal directly from internal states, removing that cost while offering a reliable way to keep the weak and strong models on similar dynamics.

The method is benchmarked on continuous diffusion language models, where it reports a new state-of-the-art on unconditional generation. On a 1.7B diffusion language model, probe guidance consistently improves multiple choice question answering benchmarks. The authors then turn the probes back on autoguidance itself, studying the setting where the strong model is a weaker checkpoint. They find that the weak model must come from a low-entropy region of training, a result presented as evidence about the actual mechanism behind autoguidance, which they describe as currently poorly understood.

The work is framed against the broader DLM context: autoregressive models generate one token per forward pass, which limits throughput and inflates latency for long sequences, while diffusion language models parallelize across positions but typically need many model evaluations to reach high quality. Probe guidance is offered both as a practical way to improve diffusion language models and as a step toward explaining how guidance works.