Abstract:We introduce discriminator guidance in the setting of Autoregressive Diffusion Models. The use of a discriminator to guide a diffusion process has previously been used for continuous diffusion models, and in this work we derive ways of using a discriminator together with a pretrained generative model in the discrete case. First, we show that using an optimal discriminator will correct the pretrained model and enable exact sampling from the underlying data distribution. Second, to account for the realistic scenario of using a sub-optimal discriminator, we derive a sequential Monte Carlo algorithm which iteratively takes the predictions from the discriminator into account during the generation process. We test these approaches on the task of generating molecular graphs and show how the discriminator improves the generative performance over using only the pretrained model.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to introduce discriminator guidance in discrete autoregressive diffusion models (ARDMs) to improve the quality of generated samples. Specifically, the author hopes to use the discriminator to correct the predictions of the pre - trained generative model, so that the generated samples are closer to the real - data distribution. ### Background and Motivation Diffusion models have received extensive attention in recent years and perform particularly well on continuous data. Kim et al. (2023) proposed a discriminator - guided method to improve the score estimation of score - based diffusion models by training a discriminator. However, this method is mainly applicable to continuous data, while ARDM deals with discrete data, so directly applying existing methods is not appropriate. ### Main Contributions 1. **Propose a discrete discriminator - guided algorithm**: - The author derived two methods of using the discriminator in ARDM: Autoregressive Discriminator Guidance (ARDG) and Sequential Monte Carlo (SMC) - based discriminator guidance. 2. **Theoretical analysis**: - The author proved that using the optimal discriminator can correct the pre - trained model and achieve accurate sampling from the real - data distribution. - For sub - optimal discriminators used in real - world situations, the author proposed the SMC algorithm, which reduces error accumulation by iteratively incorporating the discriminator's predictions into the generation process. 3. **Experimental verification**: - The author tested these methods on the molecular graph generation task, and the results show that discriminator guidance significantly improves the generation performance. ### Formula Representation - The real - data distribution is \( p_{\text{data}}(x) \), and the generative model is \( p_{\theta}(x) \). - The discriminator \( d_{\phi}(x) \) aims to distinguish between real samples and generated samples. - The optimized discriminator output is \( W_t(x_{\sigma(\leq t)})=\frac{d_{\phi}(x_{\sigma(\leq t)})}{1 - d_{\phi}(x_{\sigma(\leq t)})} \), and when the discriminator is optimal, \( W^*_t(x_{\sigma(\leq t)})=\frac{p_{\text{data}}(x_{\sigma(\leq t)})}{p_{\theta}(x_{\sigma(\leq t)})} \). ### Experimental Results The author conducted experiments on the QM9 and MOSES datasets, and the evaluation metrics include validity, uniqueness, number of atoms, molecular similarity, etc. The results show that the discriminator - guidance method significantly improves the quality of the generated samples, especially more obvious on high - dimensional data. ### Conclusion This paper successfully applies discriminator - guidance technology to discrete autoregressive diffusion models, solves the key problems in discrete - data generation, and verifies its effectiveness and superiority through experiments. This provides new ideas and methods for future research.

Discriminator Guidance for Autoregressive Diffusion Models

Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models

Classifier-Free Diffusion Guidance

Exploring Guided Sampling of Conditional GANs

Gradient Guidance for Diffusion Models: An Optimization Perspective

Theoretical Insights for Diffusion Guidance: A Case Study for Gaussian Mixture Models

Rectified Diffusion Guidance for Conditional Generation

Enhancing Diffusion-Based Image Synthesis with Robust Classifier Guidance

Classifier-Free Guidance is a Predictor-Corrector

Guided Diffusion from Self-Supervised Diffusion Features

Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models

Bridging the Gap: Addressing Discrepancies in Diffusion Model Training for Classifier-Free Guidance

Where to Diffuse, How to Diffuse, and How to Get Back: Automated Learning for Multivariate Diffusions

Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation

Guidance with Spherical Gaussian Constraint for Conditional Diffusion

Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding

Transfer Learning for Diffusion Models

The Unreasonable Effectiveness of Guidance for Diffusion Models

What does guidance do? A fine-grained analysis in a simple setting

Gradient-Free Classifier Guidance for Diffusion Model Sampling