Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
MDNS: Masked Diffusion Neural Sampler via Stochastic Optimal Control
Authors: Yuchen Zhu, Wei Guo, Jaemoo Choi, Guan-Horng Liu, Yongxin Chen, Molei Tao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the efficiency and scalability of MDNS through extensive experiments on various distributions with distinct statistical properties, where MDNS learns to accurately sample from the target distributions despite the extremely high problem dimensions and outperforms other learning-based baselines by a large margin. A comprehensive study of ablations and extensions is also provided to demonstrate the efficacy and potential of the proposed framework. |
| Researcher Affiliation | Collaboration | Yuchen Zhu1, , Wei Guo1, , Jaemoo Choi1, Guan-Horng Liu2, Yongxin Chen1, Molei Tao1 1Georgia Institute of Technology 2FAIR at Meta EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Training of Masked Diffusion Neural Sampler (MDNS) ... Algorithm 2 Sample_Trajectories: Sample trajectories and compute weights ... Algorithm 3 Training of Uniform Diffusion Neural Sampler (UDNS) ... Algorithm 4 Sample_Trajectories_Unif: Sample trajectories and compute weights for UDNS. |
| Open Source Code | Yes | Our code is available at https://github.com/yuchen-zhu-zyc/MDNS. |
| Open Datasets | No | In this section, we experimentally validate our proposed frameworks by learning to sample from the Ising model and Potts model on square lattices. At different temperatures, these models exhibit distinct behaviors, providing a rich ground for demonstrating the effectiveness of our algorithms. |
| Dataset Splits | No | The paper does not use pre-existing datasets with explicit splits. Instead, it learns to sample from statistical physics models (Ising and Potts models) where the 'data' is generated on-the-fly according to the model's parameters, and ground truth is obtained via simulation (Swendsen-Wang algorithm). |
| Hardware Specification | Yes | All experiments of learning Ising model are trained on an NVIDIA RTX A6000. ... The training of L = 4 target distributions takes around 10 minutes while the training of L = 16 target distributions takes around 20 hours (or equivalently, 4 hours on an NVIDIA RTX A100). ... All experiments of learning the Potts model are run on an NVIDIA RTX A100. |
| Software Dependencies | No | The paper mentions software components like 'Adam W optimizer', 'vision transformers (Vi T)', 'Dei T (Data-efficient image Transformers) framework', and '2-dimensional rotary position embedding', along with their respective citations. However, it does not specify version numbers for these software components or the programming environment (e.g., Python, PyTorch versions). |
| Experiment Setup | Yes | Among all the training tasks, we choose the batch size as 256, and use the Adam W optimizer [LH19] with a constant learning rate of 0.001. We always use exponential moving average (EMA) to stabilize the training, with a decay rate of 0.9999. All experiments of learning Ising model are trained on an NVIDIA RTX A6000. The training of L = 4 target distributions takes around 10 minutes while the training of L = 16 target distributions takes around 20 hours (or equivalently, 4 hours on an NVIDIA RTX A100). For 16 x 16 Ising model, we use FWDCE with a resampling frequency k = 10 and replicates R = 8 for a total of 50k iterations, among which 20k is warm-up training. |