Debiasing Guidance for Discrete Diffusion with Sequential Monte Carlo

2025-02-10

AI Summary

Existing guided sampling methods for discrete diffusion models suffer from asymptotic bias, preventing unbiased sampling from the target distribution $p_0(x_0)p(zeta|x_0)^alpha$. This work introduces sequential Monte Carlo (SMC) into discrete diffusion guidance for the first time, proposing a differentiable SMC sampler that integrates a learned unconditional diffusion process, an explicit guided transition kernel, and importance resampling. Crucially, the method provides theoretical guarantees of unbiasedness—overcoming the fundamental bias limitations of conventional classifier- and classifier-free guidance. Experiments demonstrate substantial improvements in conditional control fidelity across both image and text generation tasks. Notably, in text generation, the approach achieves superior guidance performance while maintaining low perplexity—outperforming existing guided sampling strategies.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Discrete diffusion models are a class of generative models that produce samples from an approximated data distribution within a discrete state space. Often, there is a need to target specific regions of the data distribution. Current guidance methods aim to sample from a distribution with mass proportional to $p_0(x_0) p(zeta|x_0)^alpha$ but fail to achieve this in practice. We introduce a Sequential Monte Carlo algorithm that generates unbiasedly from this target distribution, utilising the learnt unconditional and guided process. We validate our approach on low-dimensional distributions, controlled images and text generations. For text generation, our method provides strong control while maintaining low perplexity compared to guidance-based approaches.

Problem

• Debiasing guidance in discrete diffusion models
• Targeting specific data distribution regions
• Maintaining low perplexity in text generation

Innovation

• Sequential Monte Carlo algorithm
• Unbiased target distribution sampling
• Enhanced text generation control