Preprints & Publications
(In reversed chronological order. * denotes equal contribution or alphabetical order.)
2026
-
ICML 2026Rethinking the Design Space of Reinforcement Learning for Diffusion Models: On the Importance of Likelihood Estimation Beyond Loss DesignIn Forty-third International Conference on Machine Learning, 2026TL;DRShows that ELBO-based likelihood estimation is a simple yet effective design axis for RL with diffusion models. -
ICLR 2026Complexity Analysis of Normalizing Constant Estimation: from Jarzynski Equality to Annealed Importance Sampling and beyondIn The Fourteenth International Conference on Learning Representations, 2026TL;DRGives a unified complexity analysis for normalizing constant estimation methods: from Jarzynski equality to annealed importance sampling and reverse diffusion sampler. - PreprintEfficient Adjoint Matching for Fine-tuning Diffusion ModelsarXiv preprint arXiv:2605.11480, 2026TL;DRProposes an efficient adjoint matching method by reformulating SOC problem with a linear base drift and a modified terminal cost.
-
ICML 2026 SpotlightEnhancing Reasoning for Diffusion LLMs via Distribution Matching Policy OptimizationIn Forty-third International Conference on Machine Learning (Spotlight, top 2.2%), 2026TL;DRDistribution matching policy optimization via weighted denoising cross-entropy: a new RL paradigm beyond policy gradients.
2025
-
ICLR 2025Provable Benefit of Annealed Langevin Monte Carlo for Non-log-concave SamplingIn The Thirteenth International Conference on Learning Representations, 2025TL;DRProvides a theoretical guarantee of the convergence of annealed Langevin Monte Carlo from the perspective of optimal transport and Girsanov’s theorem. -
NeurIPS 2025Fast solvers for discrete diffusion models: Theory and applications of high-order algorithmsIn The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025TL;DRDevelops high-order numerical solvers that accelerate discrete diffusion models with theoretical guarantees and practical gains. -
NeurIPS 2025MDNS: Masked Diffusion Neural Sampler via Stochastic Optimal ControlIn The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025TL;DRFrames masked diffusion neural sampling as a stochastic optimal control problem for efficient discrete neural sampler training.
2024
- PreprintPlug-and-Play Controllable Generation for Discrete Masked ModelsarXiv preprint arXiv:2410.02143, 2024TL;DRDevelops a plug-and-play control method for steering masked discrete diffusion models.
2023
- Undergrad ThesisTheoretical Analysis of the Approximation Properties of Score-Based Generative ModelsUndergraduate Thesis, School of Mathematical Sciences, Peking University, 2023TL;DRStudies the convergence guarantees of score-based generative models given an imperfect score estimator and discretization errors.