Preprint · arXiv 2610.01739

Fixed-point neural samplers on discrete spaces

Introducing DIG-IN, the Discrete Gibbs Iterative Neural Sampler

Jiajun He1* Denis Blessing2* Mouyang Cheng3 Yuanqi Du4† Carles Domingo-Enrich4†
1University of Cambridge 2Karlsruhe Institute of Technology 3MIT 4Microsoft Research New England

* Equal contribution. † Equal supervision.

Overview

Sampling from a discrete distribution known only up to its normalizing constant is hard. MCMC mixes slowly across energy barriers, and existing discrete neural samplers tend to collapse onto a single mode. DIG-IN trains a masked diffusion sampler from density evaluations alone, with no training data.

Each iteration draws configurations from the current model, tempers their importance weights to a target effective sample size, and fits the model with a weighted cross-entropy. The labels in that loss are Gibbs labels: the normalized distribution over single-site changes of a configuration, a discrete analogue of the score. They give a strong learning signal and keep the sampler from collapsing onto one mode.

The DIG-IN training loop in five panels: 1, simulate and weight by revealing sites with the model and computing importance weights; 2, temper the weights with ESS control; 3, store samples, Gibbs labels and tempered weights in a replay buffer; 4, draw a minibatch and mask sites at a random time; 5, take a gradient step on the weighted soft cross-entropy with local Gibbs scores as labels.
How DIG-IN works. The target and its energy, the denoiser and its Gibbs labels, then one training iteration as in Figure 1 of the paper: data collection, importance-weight smoothing with ESS control, and a parameter update with a weighted cross-entropy whose labels are local Gibbs scores.

Alloy phase diagrams

For Ag–Pd and Cu–Au, a single sampler is conditioned on temperature T and chemical-potential difference Δμ over a DFT-fitted cluster expansion. Its free-energy estimates give the phase diagram through a common-tangent construction. Here we point into regions of that diagram and watch the sampler place atoms on the fcc cell until the structure of the phase appears.

Loading structures…
0:00 / 0:00

Solid lines are the boundaries DIG-IN predicts, dashed lines the reference. Each marker sits at the composition of the structure shown. The cell is periodic and is drawn as a 4 × 4 × 4 block of cubic fcc cells, so each site appears about twice. Drag the block to rotate it.

Amortized sampling across temperature and external field

We train a single DIG-IN model on the 50 × 50 Ising model and another on the 24 × 24 three-state Potts model, each conditioned on the inverse temperature β and an external field h. The film points at conditions on the training grid and replays how the sampler builds a configuration there, revealing one site or a small group of sites per network call.

Loading trajectories…
0:00 / 0:00

Every trajectory is a draw from the trained checkpoint, recorded call by call. Grey cells are still masked. Outlined cells were revealed in the most recent calls. Each thumbnail is one draw at its grid point; at the points the film visits, it is the draw being replayed.

Discretized Boltzmann generators

Boltzmann generators sample the equilibrium distribution of a molecule with a generative model of exact likelihood, so that importance weights remove the remaining bias. We cut every coordinate of alanine dipeptide (22 atoms) and alanine tetrapeptide (42 atoms) into bins, in internal coordinates, and train DIG-IN from the potential energy alone. A uniform point inside each bin makes a draw continuous with an exact density. The film replays draws of the trained models network call by network call and compares two million draws per run with long molecular dynamics.

Loading structures…
0:00 / 0:00

Every structure is a draw of a trained checkpoint, recorded call by call; still-masked coordinates are drawn at the network's most likely bin. The Ramachandran surfaces show ln p(φ, ψ) on 60 × 60 cells, with the three seeds pooled. Drag a molecule to rotate it.

Free energy estimation

DIG-IN's bridge learns a pair of time-reversed CTMCs between two Boltzmann distributions, from samples of the first and the energy of the second. On FEAT's alanine dipeptide tasks it carries molecular-dynamics frames from vacuum into implicit solvent (ALDP-S), and from one conformational state across a barrier into another (ALDP-T). The weights of the paths give the free-energy difference.

Loading paths…
0:00 / 0:00

The paths start from held-out MD frames and use the trained bridge of one seed. The estimates use one path per MD frame, about 250k, for each of three seeds. Drag the molecule to rotate it.

Maximum entropy reinforcement learning

In maximum entropy reinforcement learning, the improved policy at every state s is the Boltzmann distribution of the soft action value, π(a | s) ∝ exp(Q(s, a)/τ). Learning the policy is thus an amortized sampling problem over states, whose energy is a learned critic. Most policies for continuous control represent the action directly, as a Gaussian, diffusion or flow model. We instead cut every action dimension into 64 bins and use DIG-IN's masked denoiser, conditioned on the state, as the policy, trained on Gibbs labels read from the critic.

Loading the policy…
0:00 / 0:00

20 continuous control tasks of the DeepMind Control suite in MuJoCo Playground

Citation

@article{he2026fixedpoint,
  title   = {Fixed-point neural samplers on discrete spaces},
  author  = {He, Jiajun and Blessing, Denis and Cheng, Mouyang
             and Du, Yuanqi and Domingo-Enrich, Carles},
  journal = {arXiv preprint arXiv:2610.01739},
  year    = {2026}
}