The paper "Unlocking Guidance for Discrete State-Space Diffusion and Flow Models" by Hunter Nisonoff, Junhao Xiong, Stephan Allenspach, and Jennifer Listgarten introduces a general and principled method for applying guidance to discrete state-space models. The authors address the challenge of generating samples with desired properties in discrete state-spaces, which has been a significant limitation compared to continuous state-spaces where guidance approaches have been more successful. Their method, called *Discrete Guidance*, leverages continuous-time Markov processes on discrete state-spaces, making it computationally tractable to sample from a desired guided distribution. The paper begins by discussing the potential applications of generative models on discrete state-spaces, particularly in the natural sciences, and the challenges posed by discrete state-spaces. It then introduces the concept of guidance, which is crucial for conditioning the generative process on specific criteria, such as molecular conformations or protein sequences. The authors propose Discrete Guidance, which can be applied to both continuous-time diffusion and flow models on discrete state-spaces, and demonstrate its effectiveness through various experiments. Key contributions of the paper include: 1. Introducing Discrete Guidance, a general framework for applying predictor and predictor-free guidance to generative models on discrete state-spaces. 2. Developing an efficient approximation to improve sampling efficiency while maintaining sample quality. 3. Empirically evaluating the utility of Discrete Guidance across a wide range of problem domains, including small-molecule generation, DNA sequence generation, and protein sequence generation. The paper also discusses related work, compares Discrete Guidance to existing methods, and provides a detailed empirical investigation of its performance. The authors conclude by highlighting the potential of guided conditional generation in discrete state-spaces for various scientific applications and suggest future directions for research.The paper "Unlocking Guidance for Discrete State-Space Diffusion and Flow Models" by Hunter Nisonoff, Junhao Xiong, Stephan Allenspach, and Jennifer Listgarten introduces a general and principled method for applying guidance to discrete state-space models. The authors address the challenge of generating samples with desired properties in discrete state-spaces, which has been a significant limitation compared to continuous state-spaces where guidance approaches have been more successful. Their method, called *Discrete Guidance*, leverages continuous-time Markov processes on discrete state-spaces, making it computationally tractable to sample from a desired guided distribution. The paper begins by discussing the potential applications of generative models on discrete state-spaces, particularly in the natural sciences, and the challenges posed by discrete state-spaces. It then introduces the concept of guidance, which is crucial for conditioning the generative process on specific criteria, such as molecular conformations or protein sequences. The authors propose Discrete Guidance, which can be applied to both continuous-time diffusion and flow models on discrete state-spaces, and demonstrate its effectiveness through various experiments. Key contributions of the paper include: 1. Introducing Discrete Guidance, a general framework for applying predictor and predictor-free guidance to generative models on discrete state-spaces. 2. Developing an efficient approximation to improve sampling efficiency while maintaining sample quality. 3. Empirically evaluating the utility of Discrete Guidance across a wide range of problem domains, including small-molecule generation, DNA sequence generation, and protein sequence generation. The paper also discusses related work, compares Discrete Guidance to existing methods, and provides a detailed empirical investigation of its performance. The authors conclude by highlighting the potential of guided conditional generation in discrete state-spaces for various scientific applications and suggest future directions for research.