Efficient and Unbiased Sampling from Boltzmann Distributions via Variance-Tuned Diffusion Models

Authors: Fengzhe Zhang, Laurence Illing Midgley, José Miguel Hernández-Lobato

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On the DW-4, LJ-13, and alanine-dipeptide benchmarks, VT-DIS achieves effective sample sizes of approximately 80%, 35%, and 3.5%, respectively, while using only a fraction of the computational budget required by vanilla diffusion + IS or PF-ODE based IS. Our code is available at https://github.com/fz920/cov_tuned_diffusion.
Researcher Affiliation Collaboration Fengzhe Zhang EMAIL University of Cambridge Laurence I. Midgley EMAIL Ångström AI University of Cambridge José Miguel Hernández-Lobato EMAIL Ångström AI University of Cambridge
Pseudocode Yes Algorithm 1 Post-training Covariance Tuning (VT-DIS) Require: Score network sθ, initial covariances {Σϕ}, time grid {tn}N n=0, batch size M Ensure: Tuned covariances {Σϕ }
Open Source Code Yes Our code is available at https://github.com/fz920/cov_tuned_diffusion.
Open Datasets Yes DW-4 and LJ-13. We use the 107 configurations released by Klein et al. (2023). A subset of 105 configurations trains the score model and VT-DIS; the remainder forms the test set. Alanine dipeptide. We adopt the molecular dataset of Midgley et al. (2023), which provides 105 training samples and 106 test samples.
Dataset Splits Yes DW-4 and LJ-13. We use the 107 configurations released by Klein et al. (2023). A subset of 105 configurations trains the score model and VT-DIS; the remainder forms the test set. Alanine dipeptide. We adopt the molecular dataset of Midgley et al. (2023), which provides 105 training samples and 106 test samples.
Hardware Specification Yes All experiments were run on a single NVIDIA A100 GPU (80 GB); hence every GPU hour refers to wall-clock hours on that device.
Software Dependencies No The paper mentions software components like "Adam optimizer" and "cosine annealing schedule" but does not provide specific version numbers for these or other key software libraries (e.g., Python, PyTorch, CUDA).
Experiment Setup Yes GMM-2. ... Training runs for 100,000 iterations with a batch size of 1,024 for both d = 50 and d = 100. ... All models are trained for 100,000 iterations with a batch size of 512 using the Adam optimizer, a learning rate of 0.001, and a cosine annealing schedule with a minimum learning rate of 1e-6. ... VT-DIS Post-training ... Each run comprises 5,000 optimization steps with a batch size of 512 (256 for alanine dipeptide). All runs use the Adam optimizer with a learning rate of 0.01 and a cosine annealing schedule with a minimum learning rate of 1e-6.