Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction

Hongkun Dou, Zike Chen, Fengji Li, Hongjue Li, Yue Deng
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:26227-26249, 2026.

Abstract

Controllable generation with discrete diffusion models is often hindered by high computational overhead or the need for retraining. In this paper, we present Gradient-Informed Logit Correction (GILC), a plug-and-play framework that efficiently estimates guidance signals by repurposing the pretrained denoising network as a variational proxy. To circumvent the gradient instability inherent in high-dimensional discrete spaces, we introduce a Jacobian-free mechanism that directly corrects the clean prediction logits, facilitating stable and effective guidance. Our method accommodates both differentiable and non-differentiable reward functions. Extensive experiments across DNA, protein sequence, and molecular generation tasks demonstrate that GILC achieves state-of-the-art performance without additional training, frequently outperforming fine-tuning approaches.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-dou26a, title = {Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction}, author = {Dou, Hongkun and Chen, Zike and Li, Fengji and Li, Hongjue and Deng, Yue}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {26227--26249}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/dou26a/dou26a.pdf}, url = {https://proceedings.mlr.press/v306/dou26a.html}, abstract = {Controllable generation with discrete diffusion models is often hindered by high computational overhead or the need for retraining. In this paper, we present Gradient-Informed Logit Correction (GILC), a plug-and-play framework that efficiently estimates guidance signals by repurposing the pretrained denoising network as a variational proxy. To circumvent the gradient instability inherent in high-dimensional discrete spaces, we introduce a Jacobian-free mechanism that directly corrects the clean prediction logits, facilitating stable and effective guidance. Our method accommodates both differentiable and non-differentiable reward functions. Extensive experiments across DNA, protein sequence, and molecular generation tasks demonstrate that GILC achieves state-of-the-art performance without additional training, frequently outperforming fine-tuning approaches.} }
Endnote
%0 Conference Paper %T Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction %A Hongkun Dou %A Zike Chen %A Fengji Li %A Hongjue Li %A Yue Deng %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-dou26a %I PMLR %P 26227--26249 %U https://proceedings.mlr.press/v306/dou26a.html %V 306 %X Controllable generation with discrete diffusion models is often hindered by high computational overhead or the need for retraining. In this paper, we present Gradient-Informed Logit Correction (GILC), a plug-and-play framework that efficiently estimates guidance signals by repurposing the pretrained denoising network as a variational proxy. To circumvent the gradient instability inherent in high-dimensional discrete spaces, we introduce a Jacobian-free mechanism that directly corrects the clean prediction logits, facilitating stable and effective guidance. Our method accommodates both differentiable and non-differentiable reward functions. Extensive experiments across DNA, protein sequence, and molecular generation tasks demonstrate that GILC achieves state-of-the-art performance without additional training, frequently outperforming fine-tuning approaches.
APA
Dou, H., Chen, Z., Li, F., Li, H. & Deng, Y.. (2026). Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:26227-26249 Available from https://proceedings.mlr.press/v306/dou26a.html.

Related Material