Probe Guidance
A guidance method for flow-matching and diffusion models that constructs guidance signals by reusing frozen internal hidden states of an existing model, requiring no additional forward passes and eliminating extra computational overhead at inference time. It achieves state-of-the-art performance on unconditional generation tasks for continuous diffusion language models.
Timeline (last 90 days)
Probe guidance is a flow matching guidance method that constructs guidance signals by leveraging the frozen internal states of existing diffusion models.
Probe guidance extracts guidance signals directly from the frozen internal states of existing models, eliminating the need for additional forward passes during inference.
Probe guidance extracts contrastive signals from the intermediate hidden states of the strong model itself, without needing to separately train a weak model.
Probe guidance has achieved state-of-the-art (SOTA) performance on unconditional generation tasks for continuous diffusion language models.
Current publicly available information primarily focuses on probe guidance's methodology and unconditional generation benchmark results; evaluation details such as conditional generation and cross-model generalization have not yet been disclosed.
All Facts (4)
Probe guidance is a flow matching guidance method that constructs guidance signals by leveraging the frozen internal states of existing diffusion models.
50%UnverifiedProbe guidance extracts guidance signals directly from the frozen internal states of existing models, eliminating the need for additional forward passes during inference.
50%UnverifiedProbe guidance extracts contrastive signals from the intermediate hidden states of the strong model itself, without needing to separately train a weak model.
50%UnverifiedProbe guidance has achieved state-of-the-art (SOTA) performance on unconditional generation tasks for continuous diffusion language models.
50%