Continuously Tempered Diffusion Samplers

Ezra Erives, Bowen Jing, Peter Holderrieth, Tommi S. Jaakkola

Abstract

Sampling from complex, multi-modal unnormalized probability distributions is a fundamental challenge in Bayesian statistics and the physical sciences. While neural samplers have shown promise by learning transport maps, they often struggle with exploration, failing to capture isolated modes separated by high energy barriers. We present Continuously Tempered Diffusion Samplers (CTDS), a novel framework that integrates continuous tempering into the training of neural transport maps. By leveraging an augmented state space and a multi-temperature Physics-Informed Neural Network (PINN) loss, CTDS enables effective exploration of the target distribution. Our results on a challenging 40-mode Gaussian Mixture Model demonstrate that CTDS significantly outperforms existing baselines in both mode coverage and density approximation accuracy.

Introduction

The task of sampling from unnormalized probability distributions—where the density shape is known but the partition function is not—is central to many scientific domains. Neural samplers attempt to solve this by training neural networks to map a simple source distribution (e.g., a Gaussian) to a complex target. However, the training process is often hindered by poor exploration. If the target distribution contains multiple isolated modes, standard training methods frequently fail to discover all of them, leading to a sampler that lacks diversity and accuracy.

Methodology

CTDS addresses the exploration bottleneck through several technical innovations:

  1. Tempered Density Continuum: Rather than learning a static path, CTDS learns a continuum of distributions varying by time ($t$) and inverse temperature ($\beta$). High-temperature regimes "melt" the distribution, lowering energy barriers and allowing samples to transition between modes during training.
  2. Augmented Langevin Dynamics: The training process utilizes Langevin dynamics on an augmented state space $(x, \xi)$, where $\xi$ represents a continuous temperature variable. This allows the system to dynamically "heat up" to escape local minima and "cool down" to refine mode positions.
  3. Multi-Temperature PINN Loss: The model is trained using a Physics-Informed Neural Network (PINN) loss adapted for the temperature continuum, ensuring the learned vector field satisfies the continuity equation across all temperatures.
  4. Controlled Jarzynski Equality: To correct for biases introduced by the proposal distribution, the authors derive a modified Jarzynski equality. This reweighting mechanism ensures the neural network receives a high-fidelity training signal even when the proposal distribution is imperfect.

Experimental Results

The primary evaluation was conducted on a 40-mode Gaussian Mixture Model (GMM) in 2D. This synthetic benchmark serves as a rigorous "stress test" due to the extreme isolation of its modes.

Qualitative comparison of mode coverage across Baseline, NETS, and CTDS.

Quantitative Performance

CTDS demonstrates a substantial improvement over standard PINN-based samplers and Non-Equilibrium Transport Samplers (NETS). As shown in the results table, CTDS achieves a 2-Wasserstein distance ($W_2$) of 12.87, nearly halving the error of the baseline (23.90). Furthermore, the Evidence Lower Bound (ELBO) is significantly improved, reaching -0.24 compared to the baseline's -2.11.

Quantitative performance metrics on the 40-mode Gaussian Mixture Model.

Key Insights

A critical observation is the "pooling" effect: early in training, samples aggregate at high temperatures. As the model's estimate of free energy improves, samples redistribute across the temperature range. Interestingly, the version of CTDS without Jarzynski reweighting performed best in terms of mode coverage ($W_2$), suggesting that the tempered dynamics themselves provide the primary benefit for exploration.

Discussion and Limitations

While CTDS excels at mode discovery in 2D, scaling to high-dimensional spaces (e.g., protein folding) remains an open challenge. The performance of the reweighted variant is also sensitive to the accuracy of the learned free energy. Future work will focus on applying these tempered diffusion techniques to complex scientific simulations where multi-modality is a defining characteristic.

Summary updated

BibTeX
@inproceedings{erives2025continuously,
  author = {Ezra Erives and Bowen Jing and Peter Holderrieth and Tommi S. Jaakkola},
  title = {{Continuously Tempered Diffusion Samplers}},
  booktitle = {FPI @ ICLR},
  year = {2025},
  url = {https://openreview.net/forum?id=060KGPDxbH},
}
Figure