Fixed-point neural samplers on discrete spaces

Jiajun He; Denis Blessing; Mouyang Cheng; Yuanqi Du; Carles Domingo-Enrich

cs.LGOct 1, 2026

Fixed-point neural samplers on discrete spaces

Authors: Jiajun He, Denis Blessing, Mouyang Cheng, Yuanqi Du, Carles Domingo-Enrich

Organizations: MIT · Microsoft Research New England

Abstract

Sampling from discrete, unnormalized distributions without access to data is a challenging problem. Neural samplers offer a promising approach by training generative models from density evaluations directly. Despite recent progress, existing discrete neural samplers are prone to mode collapse, come without convergence guarantees when trained via fixed-point iterations, and are often tied to a specific reference process such as masked or uniform diffusion. In this work, we introduce Discrete Gibbs Iterative Neural Sampler, a fixed-point neural sampler that addresses these limitations, enabling efficient, scalable learning, substantially reducing mode collapse in practice. Our framework builds on masked diffusion and also extends to transport between pairs of distributions. We demonstrate that the resulting method scales effectively to high-dimensional systems, supports amortized sampling across different conditions, and enables accurate estimation of alloy phase diagrams.

Figures & tables

Figure 1: DIG-IN iteration. One training iteration consists of data collection, importance weight smoothing using ESS control, and parameter updates using a weighted cross-entropy loss.

Ising, $25\times 25$ , $q=2$
	disordered			critical			ordered
	$\|\Delta\ln Z\|$	$\ESS$	$\TVD_{\mathrm{mode}}$	$\|\Delta\ln Z\|$	$\ESS$	$\TVD_{\mathrm{mode}}$	$\|\Delta\ln Z\|$	$\ESS$	$\TVD_{\mathrm{mode}}$
Method	$\times 10^{3}\,{\downarrow}$	${\uparrow}$	${\downarrow}$	$\times 10^{3}\,{\downarrow}$	${\uparrow}$	${\downarrow}$	$\times 10^{3}\,{\downarrow}$	${\uparrow}$	${\downarrow}$
MDNS	$1.40_{\pm 0.79}$	$.961_{\pm.002}$	$.012_{\pm.004}$	$1.93_{\pm 1.00}$	$.925_{\pm.006}$	$.016_{\pm.006}$	$693_{\pm 0}$	$.983_{\pm.007}$	$.500_{\pm.000}$
PDNS	$2.86_{\pm 1.86}$	$.956_{\pm.001}$	$.014_{\pm.006}$	$1.47_{\pm 0.61}$	$.917_{\pm.008}$	$.016_{\pm.008}$	$693_{\pm 0}$	$.988_{\pm.003}$	$.500_{\pm.000}$
DASBS	–	–	$.007_{\pm.004}$	–	–	$.042_{\pm.028}$	–	–	$.447_{\pm.075}$

Table 1: Ising and Potts lattice spin models. We consider the Ising and Potts model at three different inverse temperatures

\beta

corresponding to disordered, critical and ordered phases. We consider three evaluation criteria: The effective sample size (

\ESS

), the error in normalization constant estimation

|\Delta\ln Z|

and the total variation distance between the empirical distribution over the

q

modes of the reference target density, denoted as

\TVD_{\mathrm{mode}}

. The table reports Mean

{\pm}

standard deviation over 3 seeds. In each column, blue shading is used to indicate methods whose pessimistic value lies within one standard deviation of the best method. Table entries with dashes indicate that the method does not support computing

\ln Z

and

\ESS

Panel of Figure 2 — Figure 2: Amortized training across $(\beta,h)$ of DIG-IN . Left: a low $\TVD_{\mathrm{mode}}$ across all $(\beta,h)$ indicates that our method avoids mode collapse and accurately learns the relative mode weights, which change under an external field $h$ . Right, top: Potts samples of the amortized model at $h=0$ in the disordered phase, near the critical point and in the ordered phase. Right, bottom: average $\ESS$ , $|\Delta\ln Z|$ and $\TVD_{\mathrm{mode}}$ over the $9\times 9$ grid of $(\beta,h)$ values, all computed with the importance weights. The results show mean $\pm$ standard deviation over three seeds.

Panel of Figure 3 — Figure 3: Alloy phase diagrams. Left: Ag–Pd ( $D=125$ , top) and Cu–Au ( $D=128$ , bottom). The grey dashed lines are the reference phase boundaries and the colored lines are the phase boundaries the model predicted. Right, top: example structures of different phases: Ag–Pd is a random fcc alloy, shown with mixed occupancy, and separates into Ag-rich and Pd-rich fcc below its miscibility gap; Cu–Au orders into Cu 3 Au, CuAu and CuAu 3 . Right, bottom: average composition error $|\Delta c|$ , $|\Delta\ln Z|/D$ and $\ESS$ over the $(T,\Delta\mu)$ values, which are the thermodynamic conditions the model was trained on. The results show mean $\pm$ standard deviation over three seeds. The best results are highlighted using blue shading.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Example configurations of the Ising and Potts model at different phases.

Systems	Ising $25\times 25$ ( $q=2$ , $D=625$ ); Potts $16\times 16$ ( $q=3$ , $D=256$ )
Phases per system	3 (disordered, critical, ordered); see Section E.1.1
Seeds	3 per (system, phase, method)
Training steps	150k
Evaluation draws	8192 per seed, independent, no MCMC correction
Backbone	periodic residual CNN, 2-D circular convolutions, 96 channels, 7 residual blocks, kernel 3, dilations covering the torus (all methods but LEAPS)
Optimizer	AdamW ( Loshchilov and Hutter, 2019 ) without weight decay (Adam for LEAPS)

Appendix

Table 2: Settings shared by every method and state point.

	Ising $25\times 25$			Potts $16\times 16$
		energy calls			energy calls
Method	train (h)	local	black box	train (h)	local	black box
DIG-IN (ours)	$6.33_{\pm 0.11}$	$1.5{\cdot}10^{6}$	$4.8{\cdot}10^{8}$	$1.89_{\pm 0.07}$	$1.5{\cdot}10^{6}$	$3.9{\cdot}10^{8}$
MDNS	$6.05_{\pm 0.09}$	$7.7{\cdot}10^{5}$	$7.7{\cdot}10^{5}$	$1.79_{\pm 0.10}$	$7.7{\cdot}10^{5}$	$7.7{\cdot}10^{5}$
PDNS	$6.42_{\pm 0.35}$	$7.7{\cdot}10^{5}$	$7.7{\cdot}10^{5}$	$2.22_{\pm 0.30}$	$7.7{\cdot}10^{5}$	$7.7{\cdot}10^{5}$
DASBS	$11.57_{\pm 0.09}$	$1.1{\cdot}10^{8}$	$6.4{\cdot}10^{10}$	$5.60_{\pm 0.28}$	$1.1{\cdot}10^{8}$	$5.3{\cdot}10^{10}$

Appendix

Table 3: Computational cost. Training time in hours (mean

{\pm}

standard deviation over 3 seeds and the three phases) and the total number of energy calls during training, with the single-site difference table counted as one call (local) or as

D(q-1)+1

calls (black box), the cost it would have if the energy were not local.

	Ising	Potts, $q=3$
Lattice	$50\times 50$ torus, $D=2500$	$24\times 24$ torus, $D=576$
Training box	$\beta\in[0.39,0.65]$ , $\|h\|\leq 0.1$	$\beta\in[0.8,1.3]$ , $\|h\|\leq 0.01$
Zero-field $\beta_{c}$	$\tfrac{1}{2}\ln(1+\sqrt{2})=0.4407$	$\ln(1+\sqrt{3})=1.00505$
Network	residual CNN on the torus: 7 blocks, 96 channels, $3\times 3$ circular convolutions, conditioned by FiLM ( Perez et al., 2018 ) on Fourier features of $(t,\beta,h)$ ; dilations $(1,1,2,4,8,1,1)$ for Ising so that 7 blocks span the $50\times 50$ torus, none for Potts
Optimizer	AdamW ( Loshchilov and Hutter, 2019 ) , learning rate $3\times 10^{-4}$ , no weight decay, gradient clipping at 1.0, EMA 0.995
Batch / buffer	1536 / 49 152	1536 / 196 608

Appendix

Table 4: Settings of the amortized runs.

	Ag–Pd	Cu–Au
Lattice, cell	fcc, $5\times 5\times 5$ primitive	fcc, $8\times 4\times 4$ primitive
Sites $D$	125	128
$\sigma_{i}=+1$	Ag	Au
Phases	one miscibility gap	Cu 3 Au (L1 2 ), CuAu (L1 0 ), CuAu 3 (L1 2 )
Training box	$T\in[200,900]$ K, $\Delta\mu\in[-0.4,0.4]$ eV	$T\in[200,900]$ K, $\Delta\mu\in[-0.4,0.4]$ eV
Evaluation $T$	200–900 K, step 25 K (29)	200–900 K, step 25 K (29)

Appendix

Table 5: The two alloy tasks.

Figure 5: Common-tangent construction on Cu–Au. Mixing free energy against composition at 200 K (left pair) and 700 K (right pair), measured from the chord through the pure compositions. In each pair: exact (left) and from the draws of one DIG-IN run (right, exact curve in grey). Dashed chords are the accepted tie lines, whose ends are the phase boundaries at that temperature.

Interpolant and sampler
Noising process	masked (absorbing state)
Reveal schedule	one site per network evaluation ( $D$ evaluations per draw)
Training labels	heat-bath conditionals from the exact single-site energy-difference table, cached with the buffered endpoint
Loss	soft cross-entropy on masked sites
Network
Backbone	periodic residual CNN, 3-D circular convolutions over the cell

Appendix

Table 6: DIG-IN hyperparameters, identical for Ag–Pd and Cu–Au.

Transport
Rate network	locally equivariant CNN: centre-masked convolution (kernel 3), then a state-dependent convolution (kernel 5); 24 channels; FiLM conditioning
Rates	Metropolis-reparametrized, clamped at 10
Free-energy head	MLP, hidden width 200, input $(t,T,\Delta\mu)$
Path discretization	$\Delta t=0.02$ (50 CTMC steps)
Horizon curriculum	from 1 step to 50, quadratic over the first 2000 optimizer steps
Prior

Appendix

Table 7: FrOGS hyperparameters, identical for Ag–Pd and Cu–Au.

Panel of Figure 6 — Figure 6: Component ablation on Ising $25\times 25$ . $\ESS$ , $|\Delta\ln Z|$ and $\TVD_{\mathrm{mode}}$ against training steps for the ordered (top), critical (middle) and disordered phase (bottom), mean over 3 seeds. Dashed: $\ln 2$ and $\TVD_{\mathrm{mode}}=0.5$ , the values of a sampler that holds only one of the two ordered modes.

	disordered, $\beta=0.30$			critical, $\beta=\beta_{c}=0.4407$			ordered, $\beta=0.60$
	$\|\Delta\ln Z\|$	$\ESS$	$\TVD_{\mathrm{mode}}$	$\|\Delta\ln Z\|$	$\ESS$	$\TVD_{\mathrm{mode}}$	$\|\Delta\ln Z\|$	$\ESS$	$\TVD_{\mathrm{mode}}$
variant	$\times 10^{3}\,{\downarrow}$	${\uparrow}$	${\downarrow}$	$\times 10^{3}\,{\downarrow}$	${\uparrow}$	${\downarrow}$	$\times 10^{3}\,{\downarrow}$	${\uparrow}$	${\downarrow}$
DIG-IN (baseline)	$1.1_{\pm 0.4}$	$.995_{\pm.000}$	$.006_{\pm.004}$	$1.6_{\pm 0.3}$	$.980_{\pm.001}$	$.015_{\pm.004}$	$0.7_{\pm 0.4}$	$.979_{\pm.011}$	$.014_{\pm.007}$
no damping	$1.2_{\pm 0.5}$	$.992_{\pm.001}$	$.010_{\pm.007}$	$1.1_{\pm 0.4}$	$.966_{\pm.007}$	$.017_{\pm.012}$	$462.0_{\pm 326.4}$	$.994_{\pm.005}$	$.336_{\pm.232}$
no Gibbs score	$2.1_{\pm 1.9}$	$.965_{\pm.001}$	$.008_{\pm.004}$	$1.5_{\pm 0.9}$	$.895_{\pm.009}$	$.011_{\pm.008}$	$26479.6_{\pm 456.9}$	$.001_{\pm.000}$	$.064_{\pm.022}$
no Gibbs, no smoothing	$1.2_{\pm 0.8}$	$.974_{\pm.002}$	$.009_{\pm.005}$	$1.0_{\pm 0.2}$	$.943_{\pm.003}$	$.019_{\pm.004}$	$693.1_{\pm 0.8}$	$.993_{\pm.000}$	$.500_{\pm.000}$

Appendix

Table 8: Component ablation, final evaluation. Mean

{\pm}

standard deviation over 3 seeds,

|\Delta\ln Z|

10^{-3}

nats. PDNS and MDNS are repeated from Table 1 .

Panel of Figure 7 — Figure 7: Bridging two Ising temperatures. Left: one row per transport, time running left to right. The map shows $P(m)$ along the path; beside it, the magnetization at $\beta_{0}$ and $\beta_{1}$ (grey: Swendsen–Wang reference, blue: the model’s draws, dashed: reweighted), and the two target modes magnified. Right, top: one draw carried along each transport, at $t=0$ , $\tfrac{1}{2}$ and $1$ (upper row: across $\beta_{c}$ , lower row: ordered). Right, bottom: $\ln Z(\beta_{1})-\ln Z(\beta_{0})$ error, $\ESS$ and reweighted $\TVD_{\mathrm{mode}}$ , mean $\pm$ standard deviation over 3 seeds.

Figure 8: Effective sample size of the amortized samplers across

(\beta,h)

. Fraction of the 5000 draws per cell, mean over seeds.

Figure 9: Effective sample size across the alloy sweep. Fraction of the 5000 draws per state point, mean over 3 seeds, for Ag–Pd (top) and Cu–Au (bottom).

Explore similar work

May 20, 2026cond-mat.stat-mech

MetaDNS: Enhancing Exploration in Discrete Neural Samplers via Well-Tempered Metadynamics

Sampling from discrete distributions with multiple modes and energy barriers is fundamental to machine learning and computational physics. Recent discrete neural samplers like MDNS suffer from mode collapse and fail to sample high-energy barrier regions between modes, which is critical for free energy estimation and understanding phase transitions. We propose Metadynamics Discrete Neural Sampler (MetaDNS), a general framework integrating well-tempered metadynamics into discrete diffusion or autoregressive samplers. By maintaining an adaptive, history-dependent bias potential along selected low-dimensional coordinates, MetaDNS forces exploration of previously inaccessible regions, enabling free energy reconstruction infeasible with standard neural samplers due to a lack of high-energy samples. On challenging low-temperature benchmarks including Ising, Potts, and the copper-gold binary alloy, MetaDNS reproduces the thermodynamic distribution. Compared to MCMC-based metadynamics, MetaDNS also achieves comparable exploration requiring fewer bias deposition steps.

Xiaochen Du, Juno Nam, Jaemoo Choi +7

Massachusetts Institute of Technology · Georgia Institute of Technology

Accelerated Sampling Molecular Dynamics Metastable Samples Phase Transitions Exploration

Ask PDF

Sep 1, 2026cond-mat.mtrl-sci

FrOGS: Discrete Neural Sampler for Independent Alloy Configurations Across Chemical Conditions

Predicting the thermodynamic properties of an alloy requires sampling its configurations across many chemical conditions and recovering free energies on a common absolute scale. Markov chain Monte Carlo (MCMC) is the standard tool, but it requires separate simulations at different conditions, and auxiliary free-energy methods such as thermodynamic integration are used to place results on a common absolute scale. Modern discrete neural samplers typically use reverse KL divergence as the objective and can be mode-seeking or biased. We present Free energy Offering Generative Sampler (FrOGS), a hybrid discrete neural sampler that couples an autoregressive model to a continuous-time Markov chain (CTMC) to be trained jointly under a single shared loss. FrOGS draws i.i.d. configurations, returns an unbiased estimate of the partition function, and gives consistent estimates of thermodynamic observables. We train a single model across a wide range of chemical conditions to produce estimates on a common absolute free-energy scale. FrOGS matches exact finite-size results on the 2D Ising model and reference phase diagrams for AgPd and CuAu, without mode collapse. We additionally compare to SEGAL, a published autoregressive baseline, and find that only FrOGS recovers the stability range of the CuAu

_3

phase.

Kyucheol Min, Elyssa Hofgard, Tess Smidt

Massachusetts Institute of Technology Cambridge, MA 02139

Markov Chain Monte Carlo Accelerated Sampling Thermodynamics Free Energy Principle Molecular Dynamics Phase Transitions+1

Ask PDF

Sep 23, 2026cs.LG

Discrete Diffusion Models via Evolving Variational Autoregressive Networks

Conventional score-based diffusion models learn scores without representing normalized densities, whereas tractable normalized models support both sampling and direct likelihood evaluation. A recent tensor-network approach provides such a representation but is largely restricted to low-dimensional lattices. Here we introduce a discrete diffusion model that parameterizes normalized probability distributions using variational autoregressive networks. Explicit Markov jump operators govern the forward noising and reverse denoising dynamics, extending discrete diffusion models with normalized distributions to spin systems on higher-dimensional lattices. We apply this framework to the two- and three-dimensional Ising models across ordered, critical, and disordered regimes, accurately computing thermodynamic quantities including free energy, energy, and magnetization. We further integrate the framework with Monte Carlo sampling, using adaptive diffusion steps to maintain high acceptance rates even at low temperatures while enhancing sample diversity. These results establish a neural-network framework for the discrete diffusion model with normalized probability distributions.

Kewen Pan, Ying Tang

Institute of Fundamental and Frontier Sciences, University of Electronic Science and Technology of China, Chengdu 611731, China · School of Physics, University of Electronic Science and Technology of China, Chengdu 611731, China · Key Laboratory of Quantum Physics and Photonic Quantum Information, Ministry of Education, University of Electronic Science and Technology of China, Chengdu 611731, China +1

Score-Based Diffusion Model Spatial Photonic Ising Machines Diffusion Models Tensor Networks Conditional Variational Autoencoder

Ask PDF