pith. sign in

arxiv: 2605.21722 · v1 · pith:B3VXWJHOnew · submitted 2026-05-20 · ❄️ cond-mat.stat-mech · cond-mat.mtrl-sci· cs.LG

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

Xiaochen Du , Juno Nam , Jaemoo Choi , Wei Guo , Sathya Edamadaka , Junyi Sha , Elton Pan , Yongxin Chen
show 2 more authors
Molei Tao Rafael G\'omez-Bombarelli
This is my paper

Pith reviewed 2026-05-22 08:05 UTC · model grok-4.3

classification ❄️ cond-mat.stat-mech cond-mat.mtrl-scics.LG
keywords metadynamicsdiscrete neural samplersfree energy estimationIsing modelPotts modelmode collapsediffusion modelsenergy barriers
0
0 comments X

The pith

MetaDNS adds an adaptive bias potential to discrete neural samplers so they can cross high-energy barriers between modes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MetaDNS as a way to integrate well-tempered metadynamics into discrete diffusion or autoregressive samplers. It maintains a history-dependent bias along chosen low-dimensional coordinates that pushes the sampler into regions that standard neural methods avoid because of mode collapse. A reader would care because those missing high-energy samples prevent accurate free-energy reconstruction and the study of phase transitions in systems such as magnets and alloys. The method is shown to recover the correct thermodynamic distribution on low-temperature Ising, Potts, and binary-alloy benchmarks while using fewer bias steps than MCMC-based metadynamics.

Core claim

MetaDNS maintains an adaptive, history-dependent bias potential along selected low-dimensional coordinates inside discrete neural samplers. This bias forces the sampler to visit previously inaccessible high-energy barrier regions, thereby enabling free-energy reconstruction from the generated distribution that standard neural samplers cannot perform because they lack sufficient high-energy samples. On the Ising, Potts, and copper-gold alloy models at low temperature, MetaDNS reproduces the thermodynamic distribution and reaches comparable exploration to MCMC metadynamics with fewer bias-deposition steps.

What carries the argument

An adaptive, history-dependent bias potential deposited along chosen low-dimensional collective variables, which gradually flattens energy barriers inside the discrete sampler.

If this is right

  • Standard neural samplers can now generate the high-energy samples required for free-energy estimation in multimodal discrete systems.
  • Thermodynamic distributions are recovered on low-temperature Ising, Potts, and copper-gold alloy models.
  • Comparable exploration quality to MCMC metadynamics is reached with fewer bias-deposition steps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bias mechanism could be applied to other discrete generative tasks where rare events or phase coexistence matter.
  • Extending the coordinate choice to learned collective variables might further reduce the number of required bias steps.
  • The framework may allow larger lattice sizes in simulations of phase transitions by improving sampling efficiency.

Load-bearing premise

The chosen low-dimensional coordinates are sufficient to capture the relevant collective variables and energy barriers for effective bias deposition in the discrete setting.

What would settle it

If MetaDNS samples from the low-temperature Ising model fail to produce configurations across the known energy barrier and the reconstructed free energy deviates from the exact value by more than statistical error, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.21722 by Elton Pan, Jaemoo Choi, Juno Nam, Junyi Sha, Molei Tao, Rafael G\'omez-Bombarelli, Sathya Edamadaka, Wei Guo, Xiaochen Du, Yongxin Chen.

Figure 1
Figure 1. Figure 1: Scheme for MetaDNS. (Left) State of the art (SOTA) discrete neural samplers suffer from mode collapse at low temperatures: samples concentrate in a single deep minimum of the multi-modal energy landscape. (Middle) MetaDNS adds a history-dependent bias potential (purple) that raises the energy in frequently visited regions, flattening the landscape and encouraging exploration across energy barriers to visit… view at source ↗
Figure 2
Figure 2. Figure 2: Ising model results at L = 16: (a-c) Up-spin concen￾tration distributions at three temperatures (β = 0.6, 0.44, 0.28). MetaDNS (ours) captures the full bimodal distribution at low tem￾perature while MDNS suffers from mode collapse. (d-f) Free energy per site versus inverse temperature for x↑ = 0.25, 0.5, and 0.75. Lack of samples at intermediate spin concentrations prevents MDNS from estimating free energi… view at source ↗
Figure 3
Figure 3. Figure 3: Collective variable (CV) distributions for Potts (q = 3, L = 16) models at two temperatures: (a) low and (b) critical. 2D CV space (CV 1 vs. CV 2) comparing SW (ground truth), MDNS, MetaDNS (ours), and MCMC-based WT-MetaD. At low and critical temperatures, MDNS exhibits mode collapse, discovering only one of three modes. MetaDNS successfully covers all modes and interspace regions, matching WT-MetaD covera… view at source ↗
Figure 4
Figure 4. Figure 4: Free energy profiles and convergence for Potts (q = 3) at L = 16 comparing MetaDNS (ours) with MCMC-based WT￾MetaD. (a-c) Free energy profiles along CV 1 at high (β = 0.5), critical (β = 1.01), and low (β = 1.2) temperatures show￾ing agreement with WT-MetaD. (d-f) Convergence speedup of MetaDNS over WT-MetaD at the same temperatures (RMSE vs. steps). fers from high correlation between sequential MC steps a… view at source ↗
Figure 5
Figure 5. Figure 5: Cu-Au alloy results at 4 × 4 × 4. (a-b) Crystal struc￾tures of Cu3Au and CuAu ordered phases. (c-e) Au concentration distributions at 500K, 680K, and 1200K showing MDNS mode collapse at 500K, capturing only the xAu = 0.5 mode, whereas MetaDNS (ours) matches the ground truth. (f) Free energy profiles at 500K showing MetaDNS (ours) agreement with WT-MetaD. (g) Convergence speedup of MetaDNS over WT-MetaD at … view at source ↗
Figure 6
Figure 6. Figure 6: L = 16 Ising model free energy per site and deviation from WT-MetaD reference for x↑ = (a) 0.25, (b) 0.5, and (c) 0.75. MetaDNS (ours) is able to obtain samples across the full composition range at low β = 0.6 and critical β = 0.4407 temperatures for free energy (mixing energy) estimation but MDNS fails at low temperature. a d b c e f [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ising model results at L = 8. (a-c) Up-spin concentration distributions across three temperatures, showing excellent agreement between MetaDNS (ours) and SW ground truth across all temperatures after reweighting but MDNS fails to sample the full down-spin mode at low temperature. (d-f) Free energy per site for x↑ = 0.25, 0.5, and 0.75 compared with WT-MetaD reference and deviation from WT-MetaD reference. … view at source ↗
Figure 8
Figure 8. Figure 8: Two-point correlation functions for Ising models at L ∈ {(a-c) 4, (d-f) 8, and (g-i) 16} and three temperatures. MetaDNS (ours, after reweighting) and MDNS both show strong agreement with SW ground truth across all conditions, validating that MetaDNS (ours) maintains correct statistical properties while achieving improved mode exploration. (a) MDNS 50k steps (b) MDNS 100k steps (c) MetaDNS [PITH_FULL_IMAG… view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparison of L = 16 Ising model samples at low temperature. (a) MDNS shows mode collapse when trained for 50k steps, sampling predominantly from a single spin configuration. (b) Mode collapses persists even when MDNS was trained for 100k steps (2x original). (c) MetaDNS (ours) successfully discovers diverse configurations with well-formed domains of both spin orientations, demonstrating effective m… view at source ↗
Figure 10
Figure 10. Figure 10: Visual comparison of L = 16 Ising model samples at additional low temperature β > βcrit. (a) β = 0.5, (b) β = 0.7, and (c) β = 0.8. Regardless of exact temperature value, MDNS shows mode collapse. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: CV distributions for Potts (q = 3) models at L = 16 and high temperature. MDNS agrees with SW ground truth for this temperature, favoring the random configurations with CV ≈ (0, 0). MetaDNS (ours) and WT-MetaD are able to sample the CV landscape more broadly. 0 1 CV 1 (x) 1 0 1 CV 2 (y) SW 0 1 CV 1 (x) 1 0 1 MDNS 0 1 CV 1 (x) 1 0 1 MDNS (anneal) 0 1 CV 1 (x) 1 0 1 MetaDNS 0 1 CV 1 (x) 1 0 1 WT-MetaD 0 50 … view at source ↗
Figure 12
Figure 12. Figure 12: Mode collapse comparison for MDNS L = 16 Potts model (center) at β = 1.2 even when warm-started at high temperature. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: CV distributions for Potts (q = 3) models at L = 8 and three temperatures: (a) low β = 1.2, (b) critical β = 1.005, and (c) high β = 0.5. Similar to the L = 16 case, MetaDNS (ours) and WT-MetaD are able to sample the CV landscape more broadly at all temperatures, with all modes present. MDNS, however, exhibits mode collapse at low temperature β = 1.2 and only samples one of the three modes present in SW g… view at source ↗
Figure 14
Figure 14. Figure 14: Two-point correlation functions for Potts (q = 3) models at L ∈ {(a-c) 4, (d-f) 8, and (g-i) 16} and three temperatures. MetaDNS (ours, after reweighting) and MDNS both show strong agreement with SW ground truth across all conditions, validating that MetaDNS (ours) maintains correct statistical properties while achieving improved mode exploration. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visual comparison of L = 16 Potts (q = 3) model samples at low temperature. (a) MDNS model 1 shows mode collapse, sampling predominantly from a single phase configuration; (b) MDNS model 2 exhibits similar mode collapse in a different phase; (c) MDNS still suffers from mode collapse after warm-starting from high temperature; (d) MetaDNS (ours) successfully discovers diverse configurations with well-formed… view at source ↗
Figure 16
Figure 16. Figure 16: Learning dynamics for MetaDNS vs. MCMC-based WT-MetaD for the first 50k steps. WT-MetaD remains concentrated near random configurations (CV ≈ (0, 0)) before gradually spreading to the target modes, whereas MetaDNS rapidly discovers and resolves the distinct low-energy basins. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Sensitivity of MetaDNS to hyperparameters for the L = 16 Ising model at β = 0.6. σ ∈ {0.01, 0.03, 0.05}, h ∈ {0.1kBT, 0.5kBT}, γ ∈ {5, 10}, and number of CV bins ∈ {129, 257}. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Up-spin concentration distributions sensitivity analysis for the L = 16 Ising model at β = 0.6 with 129 bins. 0 2 4 6 8 10 Density = 0.01, h=0.1 kBT, = 5 = 0.01, h=0.1 kBT, = 10 = 0.01, h=0.5 kBT, = 5 = 0.01, h=0.5 kBT, = 10 0 2 4 6 8 10 Density = 0.03, h=0.1 kBT, = 5 = 0.03, h=0.1 kBT, = 10 = 0.03, h=0.5 kBT, = 5 = 0.03, h=0.5 kBT, = 10 0.0 0.5 1.0 Spin Concentration 0 2 4 6 8 10 Density = 0.05, h=0.1 kB… view at source ↗
Figure 19
Figure 19. Figure 19: Up-spin concentration distributions sensitivity analysis for the L = 16 Ising model at β = 0.6 with 257 bins. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Sensitivity of MetaDNS to hyperparameters for the L = 16 Potts model at β = 1.2. σ ∈ {0.01, 0.03, 0.05}, h ∈ {0.1kBT, 0.5kBT},γ ∈ {5, 10}. 1.0 0.5 0.0 0.5 1.0 CV 2 (y) = 0.01, h=0.1 kBT, = 5 = 0.01, h=0.1 kBT, = 10 = 0.01, h=0.5 kBT, = 5 = 0.01, h=0.5 kBT, = 10 1.0 0.5 0.0 0.5 1.0 CV 2 (y) = 0.03, h=0.1 kBT, = 5 = 0.03, h=0.1 kBT, = 10 = 0.03, h=0.5 kBT, = 5 = 0.03, h=0.5 kBT, = 10 0 1 CV 1 (x) 1.0 0.5 0.… view at source ↗
Figure 21
Figure 21. Figure 21: CV distributions sensitivity analysis for Potts (q = 3) models at L = 16 and low temperature. Mode collapse is observed only for the σ = 0.01, h = 0.1kBT, γ = 5 case. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Sensitivity of MetaDNS to hyperparameters for the 4 × 4 × 4 Cu-Au alloy at 500K. σ ∈ {0.01, 0.03, 0.05}, h ∈ {0.1kBT, 0.5kBT},γ ∈ {5, 10}. 0 2 4 6 8 10 Density = 0.01, h=0.1 kBT, = 5 = 0.01, h=0.1 kBT, = 10 = 0.01, h=0.5 kBT, = 5 = 0.01, h=0.5 kBT, = 10 0 2 4 6 8 10 Density = 0.03, h=0.1 kBT, = 5 = 0.03, h=0.1 kBT, = 10 = 0.03, h=0.5 kBT, = 5 = 0.03, h=0.5 kBT, = 10 0.25 0.50 0.75 Au Concentration 0 2 4 6… view at source ↗
Figure 23
Figure 23. Figure 23: Au concentration distributions sensitivity analysis for the 4 × 4 × 4 Cu-Au alloy at 500K. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_23.png] view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces MetaDNS, a framework integrating well-tempered metadynamics into discrete neural samplers (diffusion or autoregressive) via an adaptive, history-dependent bias potential deposited along selected low-dimensional collective variables. This is claimed to overcome mode collapse and enable sampling of high-energy barrier regions in discrete distributions, allowing free-energy reconstruction on low-temperature Ising, Potts, and copper-gold alloy benchmarks, while achieving comparable exploration to MCMC metadynamics with fewer bias deposition steps.

Significance. If the bias integration preserves the correct stationary distribution and the empirical reproduction of thermodynamic quantities is quantitatively validated, the approach could meaningfully extend neural samplers to multimodal discrete systems with barriers, offering efficiency gains for free-energy estimation in statistical mechanics and related machine-learning tasks.

major comments (2)
  1. Abstract: the claim that MetaDNS 'reproduces the thermodynamic distribution' on Ising, Potts, and alloy benchmarks is unsupported by any quantitative metrics, error bars, overlap measures, or validation details, making it impossible to assess whether the sampled distributions match the target Boltzmann weights within statistical error.
  2. Methods (bias incorporation): the mapping from the continuous, history-dependent bias potential (deposited on low-dimensional CVs) to the discrete-state probabilities or logits of the neural sampler is not shown to guarantee that the effective stationary distribution equals the original energy plus bias. If the bias is applied only approximately (e.g., via grid interpolation or post-hoc reweighting), both the exploration guarantees and the subsequent free-energy reconstruction are undermined.
minor comments (1)
  1. Clarify the precise criterion used to select the low-dimensional coordinates as collective variables and demonstrate that they capture the relevant energy barriers for the chosen discrete models.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and constructive comments on our manuscript. We address each major comment point by point below, providing clarifications and indicating the revisions made to strengthen the presentation and theoretical grounding of MetaDNS.

read point-by-point responses

  1. Referee: Abstract: the claim that MetaDNS 'reproduces the thermodynamic distribution' on Ising, Potts, and alloy benchmarks is unsupported by any quantitative metrics, error bars, overlap measures, or validation details, making it impossible to assess whether the sampled distributions match the target Boltzmann weights within statistical error.

    Authors: We agree that the original abstract statement would be strengthened by explicit quantitative support. In the revised manuscript we have updated the abstract to reference the specific validation metrics reported in the Results section, including KL divergence, total variation distance, and mean absolute deviations in reconstructed free energies, each accompanied by error bars obtained from multiple independent sampling runs. These metrics confirm agreement with the target Boltzmann distribution within statistical error on the Ising, Potts, and copper-gold alloy benchmarks. revision: yes

  2. Referee: Methods (bias incorporation): the mapping from the continuous, history-dependent bias potential (deposited on low-dimensional CVs) to the discrete-state probabilities or logits of the neural sampler is not shown to guarantee that the effective stationary distribution equals the original energy plus bias. If the bias is applied only approximately (e.g., via grid interpolation or post-hoc reweighting), both the exploration guarantees and the subsequent free-energy reconstruction are undermined.

    Authors: We acknowledge the need for an explicit guarantee. The original Methods section describes direct incorporation of the bias by redefining the energy fed to the neural sampler as E_biased(s) = E(s) + V(CV(s)), where V is the well-tempered metadynamics bias evaluated on the chosen collective variables. Because the discrete neural sampler (diffusion or autoregressive) is trained or conditioned to target the Boltzmann distribution of E_biased, the stationary distribution is the desired biased distribution by construction; no grid interpolation or post-hoc reweighting is employed during sampling. We have added a dedicated paragraph in the revised Methods that derives this stationarity preservation and clarifies the subsequent reweighting procedure used for free-energy reconstruction. revision: yes

Circularity Check

0 steps flagged

No circularity: method combines established techniques with empirical validation

full rationale

The paper describes a framework that integrates well-tempered metadynamics into discrete diffusion or autoregressive samplers by adding a history-dependent bias potential along chosen low-dimensional coordinates. This construction is presented as a direct methodological extension rather than a derivation in which any central quantity (such as the effective sampling distribution or free-energy estimate) is defined in terms of itself or recovered by construction from fitted parameters. No load-bearing step reduces to a self-citation chain, an ansatz smuggled via prior work, or a fitted input relabeled as a prediction; the reported performance on Ising, Potts, and alloy benchmarks is obtained by running the combined sampler and is therefore falsifiable against external MCMC references. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; details on parameters and assumptions are limited. Potential free parameters exist in bias construction and coordinate choice.

free parameters (1)
  • bias deposition rate and width
    Well-tempered metadynamics typically requires tunable parameters for bias height and Gaussian width that may be fitted or chosen per problem.
axioms (1)
  • domain assumption Low-dimensional coordinates can serve as effective collective variables to represent barriers in discrete configuration spaces.
    Invoked when applying metadynamics bias along selected coordinates in the discrete sampler.

pith-pipeline@v0.9.0 · 5725 in / 1249 out tokens · 46242 ms · 2026-05-22T08:05:04.030714+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

  1. [1]

    , year =

    Glauber, Roy J. , year =. Time-. Journal of Mathematical Physics , volume =

    work page

  2. [2]

    2010 , month = may, journal =

    Metadynamics Convergence Law in a Multidimensional System , author =. 2010 , month = may, journal =

    work page 2010

  3. [3]

    Reconstructing the

    Micheletti, Cristian and Laio, Alessandro and Parrinello, Michele , year =. Reconstructing the. Physical Review Letters , volume =

    work page

  4. [4]

    Barducci, Alessandro and Bussi, Giovanni and Parrinello, Michele , year =. Well-. Physical Review Letters , volume =

    work page

  5. [5]

    2023 , journal =

    RoFormer: Enhanced Transformer with Rotary Position Embedding , author =. 2023 , journal =

    work page 2023

  6. [6]

    and Parrinello, Michele and Voth, Gregory A

    Dama, James F. and Parrinello, Michele and Voth, Gregory A. , year =. Well-. Physical Review Letters , volume =

    work page

  7. [7]

    and Wang, Jian-Sheng , year =

    Swendsen, Robert H. and Wang, Jian-Sheng , year =. Nonuniversal Critical Dynamics in. Physical Review Letters , volume =

    work page

  8. [8]

    Journal of Physics: Condensed Matter , volume =

    Chang, Jin Hyun and Kleiven, David and Melander, Marko and Akola, Jaakko and Garcia-Lastra, Juan Maria and Vegge, Tejs , year =. Journal of Physics: Condensed Matter , volume =

    work page

  9. [9]

    2022 , month = apr, journal =

    Sampling Lattices in Semi-Grand Canonical Ensemble with Autoregressive Machine Learning , author =. 2022 , month = apr, journal =

    work page 2022

  10. [10]

    2024 , month = feb, journal =

    Sampling Algorithms in Statistical Physics: A Guide for Statistics and Machine Learning , author =. 2024 , month = feb, journal =

    work page 2024

  11. [11]

    1953 , month = jun, journal =

    Equation of State Calculations by Fast Computing Machines , author =. 1953 , month = jun, journal =

    work page 1953

  12. [12]

    Proximal

    Guo, Wei and Choi, Jaemoo and Zhu, Yuchen and Tao, Molei and Chen, Yongxin , year =. Proximal. International Conference on Learning Representations , url =

    work page

  13. [13]

    and Jaakkola, Tommi , year =

    Holderrieth, Peter and Albergo, Michael S. and Jaakkola, Tommi , year =. Proceedings of the 42nd International Conference on Machine Learning , publisher =

    work page

  14. [14]

    2002 , journal =

    Escaping Free-Energy Minima , author =. 2002 , journal =

    work page 2002

  15. [15]

    , year =

    Liu, Sulin and Ramadge, Peter and Adams, Ryan P. , year =. Generative. Proceedings of the 41st

    work page

  16. [16]

    2019 , month = may, journal =

    work page 2019

  17. [17]

    2018 , month = jul, journal =

    First-Principles Statistical Mechanics of Multicomponent Crystals , author =. 2018 , month = jul, journal =

    work page 2018

  18. [18]

    Enhancing

    Nam, Juno and M. Enhancing. 2026 , booktitle =

    work page 2026

  19. [19]

    2020 , month = feb, journal =

    Asymptotically Unbiased Estimation of Physical Observables with Neural Samplers , author =. 2020 , month = feb, journal =

    work page 2020

  20. [20]

    Discrete

    Ou, Zijing and Zhang, Ruixiang and Li, Yingzhen , year =. Discrete. Advances in Neural Information Processing Systems , volume =

    work page

  21. [21]

    Ou, Jingyang and Nie, Shen and Xue, Kaiwen and Zhu, Fengqi and Sun, Jiacheng and Li, Zhenguo and Li, Chongxuan , year =. Your. International Conference on Learning Representations , url =

    work page

  22. [22]

    2014 , month = sep, journal =

    Effective Ergodicity in Single-Spin-Flip Dynamics , author =. 2014 , month = sep, journal =

    work page 2014

  23. [23]

    Wu, Dian and Wang, Lei and Zhang, Pan , year =. Solving. Physical Review Letters , volume =

    work page

  24. [24]

    Advances in Neural Information Processing Systems , volume =

    Zhu, Yuchen and Guo, Wei and Choi, Jaemoo and Liu, Guan-Horng and Chen, Yongxin and Tao, Molei , year =. Advances in Neural Information Processing Systems , volume =

    work page

  25. [25]

    2019 , journal =

    Boltzmann Generators: Sampling Equilibrium States of Many-Body Systems with Deep Learning , author =. 2019 , journal =

    work page 2019

  26. [26]

    2024 , booktitle =

    Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution , author =. 2024 , booktitle =

    work page 2024

  27. [27]

    2022 , booktitle =

    Path Integral Sampler: A Stochastic Control Approach for Sampling , author =. 2022 , booktitle =

    work page 2022

  28. [28]

    2021 , booktitle =

    Score-Based Generative Modeling through Stochastic Differential Equations , author =. 2021 , booktitle =

    work page 2021

  29. [29]

    2025 , url =

    Sampling from Energy Distributions with Target Concrete Score Identity , author =. 2025 , url =

    work page 2025

  30. [30]

    Exchange

    Hukushima, Koji and Nemoto, Koji , year =. Exchange. Journal of the Physical Society of Japan , volume =

    work page

  31. [31]

    2024 , journal =

    Improving Control Based Importance Sampling Strategies for Metastable Diffusions via Adapted Metadynamics , author =. 2024 , journal =

    work page 2024

  32. [32]

    Du, Xiaochen and Nam, Juno and Liu, Sulin and. Scaling. 2026 , month = mar, url =

    work page 2026

  33. [33]

    Targeted

    Zhang, Jun and Yang, Yi Isaac and No. Targeted. 2019 , month = oct, journal =

    work page 2019

  34. [34]

    Liu, Guan-Horng and Choi, Jaemoo and Chen, Yongxin and Miller, Benjamin Kurt and Chen, Ricky T. Q. , year =. Adjoint. Advances in Neural Information Processing Systems , volume =

    work page

  35. [35]

    Discrete

    Guo, Wei and Zhu, Yuchen and Du, Xiaochen and Nam, Juno and Chen, Yongxin and G. Discrete. 2026 , booktitle =

    work page 2026

  36. [36]

    Enhanced

    Zhu, Kai and Trizio, Enrico and Zhang, Jintu and Hu, Renling and Jiang, Linlong and Hou, Tingjun and Bonati, Luigi , year =. Enhanced. Chemical Reviews , volume =

    work page