pith. machine review for the scientific record. sign in

arxiv: 2605.00229 · v1 · submitted 2026-04-30 · 📊 stat.ML · cs.LG· math.OC

Recognition: unknown

A unified perspective on fine-tuning and sampling with diffusion and flow models

Carles Domingo-Enrich , Yuanqi Du , Michael S. Albergo
Authors on Pith no claims yet

Pith reviewed 2026-05-09 19:40 UTC · model grok-4.3

classification 📊 stat.ML cs.LGmath.OC
keywords diffusion modelsflow modelsfine-tuningexponential tiltingscore matchingadjoint methodsstochastic optimal controlunnormalized sampling
0
0 comments X

The pith

Exponential tilting unifies sampling from unnormalized densities with reward fine-tuning of diffusion and flow models.

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

The paper studies training diffusion and flow models to sample from target distributions formed by exponentially tilting a base density. This single formulation covers both drawing samples from unnormalized probabilities and fine-tuning pre-trained models with reward signals. The authors build a unified framework that merges stochastic optimal control methods with non-equilibrium thermodynamics views. They then decompose gradients to show that adjoint matching and novel score matching keep finite variance while other score matching variants do not, establish norm bounds for the lean adjoint ODE, and adapt existing loss functions with new Crooks and Jarzynski identities. These results matter because they clarify why certain training procedures remain stable when applied to high-dimensional tasks such as image generation.

Core claim

By recasting the problem as sampling under exponential tilting of a base density, the authors create a single perspective that includes both unnormalized sampling and reward fine-tuning of diffusion and flow models. Within this view they obtain bias-variance decompositions proving finite gradient variance for Adjoint Matching and Novel Score Matching, unlike Target and Conditional Score Matching. They further supply norm bounds on the lean adjoint ODE and adapt the CMCD and NETS losses while deriving novel Crooks and Jarzynski identities for the tilting case. Experiments on reward fine-tuning of Stable Diffusion 1.5 and 3 confirm that the theoretical distinctions appear in practice.

What carries the argument

Exponential tilting of a base density, analyzed jointly through stochastic optimal control adjoint methods and score matching techniques.

If this is right

  • Adjoint Matching and Novel Score Matching possess finite gradient variance, allowing stable optimization where other score matching losses diverge.
  • Norm bounds on the lean adjoint ODE supply theoretical justification for the observed reliability of adjoint-based training.
  • Adapted CMCD and NETS losses together with the new Crooks and Jarzynski identities extend directly to the exponential tilting setting.
  • The same distinctions between methods apply to both unnormalized sampling and reward fine-tuning tasks.

Where Pith is reading between the lines

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

  • The finite-variance results could guide choice of training objectives when applying similar control ideas to other generative architectures.
  • The thermodynamic identities may suggest new sampling algorithms outside the diffusion setting.
  • Practical success on Stable Diffusion indicates the framework may scale to larger models, provided high-dimensional instabilities are monitored.
  • The unification invites direct comparisons between adjoint and score-based methods on shared benchmark tasks beyond image generation.

Load-bearing premise

The exponential tilting formulation captures practical fine-tuning goals without introducing instabilities or excessive computation costs in high-dimensional settings.

What would settle it

An experiment that measures infinite gradient variance for Adjoint Matching under the exponential tilting objective during reward fine-tuning of a diffusion model on image data.

Figures

Figures reproduced from arXiv: 2605.00229 by Carles Domingo-Enrich, Michael S. Albergo, Yuanqi Du.

Figure 1
Figure 1. Figure 1: Trade-offs between per-prompt diversity (DreamSim variance, [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Trade-offs between per-prompt diversity (DreamSim variance, [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quality metrics for the base Stable Diffusion 3 model and models fine-tuned at [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
read the original abstract

We study the problem of training diffusion and flow generative models to sample from target distributions defined by an exponential tilting of a base density; a formulation that subsumes both sampling from unnormalized densities and reward fine-tuning of pre-trained models. This problem can be approached from a stochastic optimal control (SOC) perspective, using adjoint-based or score matching methods, or from a non-equilibrium thermodynamics perspective. We provide a unified framework encompassing these approaches and make three main contributions: (i) bias-variance decompositions revealing that Adjoint Matching/Sampling and Novel Score Matching have finite gradient variance, while Target and Conditional Score Matching do not; (ii) norm bounds on the lean adjoint ODE that theoretically support the effectiveness of adjoint-based methods; and (iii) adaptations of the CMCD and NETS loss functions, along with novel Crooks and Jarzynski identities, to the exponential tilting setting. We validate our analysis with reward fine-tuning experiments on Stable Diffusion 1.5 and 3.

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 / 3 minor

Summary. The manuscript develops a unified framework for training diffusion and flow models to sample from target distributions obtained by exponential tilting of a base density. This formulation covers both sampling from unnormalized densities and reward-based fine-tuning of pretrained models. The work bridges stochastic optimal control (SOC) methods (adjoint-based and score-matching) with non-equilibrium thermodynamics perspectives, deriving (i) bias-variance decompositions showing finite gradient variance for Adjoint Matching/Sampling and Novel Score Matching but not for Target/Conditional Score Matching, (ii) norm bounds on the lean adjoint ODE, and (iii) adaptations of the CMCD and NETS losses together with novel Crooks and Jarzynski identities in the tilting setting. These are validated through reward fine-tuning experiments on Stable Diffusion 1.5 and 3.

Significance. If the derivations hold, the paper supplies concrete theoretical explanations for the empirical behavior of adjoint and score-matching estimators under tilting, including why certain methods exhibit finite variance while others do not. The norm bounds on the lean adjoint ODE provide a rigorous basis for the stability of adjoint-based fine-tuning, and the adapted thermodynamic identities extend classical fluctuation theorems to the generative-model setting. The experiments on Stable Diffusion directly test the framework in a high-dimensional practical regime, lending credibility to the claims. Credit is due for the explicit bias-variance decompositions, the parameter-free character of the bounds, and the direct experimental validation.

major comments (2)
  1. [Section 4 (Bias-Variance Analysis)] The bias-variance decompositions in contribution (i) are central to the claim that Adjoint Matching/Sampling and Novel Score Matching are preferable; the manuscript should explicitly state the assumptions on the tilting function and the base density under which the finite-variance result holds, and verify that these assumptions are satisfied by the reward functions used in the Stable Diffusion experiments.
  2. [Section 5 (Adjoint ODE Analysis)] The norm bounds on the lean adjoint ODE (contribution (ii)) are load-bearing for the theoretical support of adjoint methods; the proof should clarify whether the bounds remain uniform in the dimension of the data space, as high-dimensional image models (e.g., Stable Diffusion) could otherwise render the constants prohibitive.
minor comments (3)
  1. [Section 2] The notation for the exponential tilting parameter and the base density should be introduced once and used consistently; occasional redefinition in later sections obscures the connection between the SOC and thermodynamics viewpoints.
  2. [Section 6] Figure captions for the Stable Diffusion fine-tuning results should include the precise reward function, number of fine-tuning steps, and baseline methods compared, to allow direct replication of the reported improvements.
  3. [Section 6] A brief discussion of computational overhead (wall-clock time or memory) for the adapted CMCD/NETS losses versus standard score matching would strengthen the practical takeaway.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their positive evaluation, detailed summary, and constructive major comments. We appreciate the recognition of the bias-variance decompositions, norm bounds, and thermodynamic identities. We address each comment below and will incorporate clarifications in the revised manuscript.

read point-by-point responses

  1. Referee: [Section 4 (Bias-Variance Analysis)] The bias-variance decompositions in contribution (i) are central to the claim that Adjoint Matching/Sampling and Novel Score Matching are preferable; the manuscript should explicitly state the assumptions on the tilting function and the base density under which the finite-variance result holds, and verify that these assumptions are satisfied by the reward functions used in the Stable Diffusion experiments.

    Authors: We agree that an explicit statement of assumptions will strengthen the presentation. The finite-variance results for Adjoint Matching/Sampling and Novel Score Matching hold under the assumptions that the tilting function is bounded (or has bounded gradients) and the base density has finite second moments; these are standard regularity conditions ensuring the relevant expectations exist and the variance is controlled. In the Stable Diffusion experiments the reward functions are normalized and clipped to [0,1] (as is common in reward fine-tuning), which satisfies boundedness. We will add a short paragraph at the beginning of Section 4 stating these assumptions together with a verification sentence confirming they hold for the reported experiments. revision: yes

  2. Referee: [Section 5 (Adjoint ODE Analysis)] The norm bounds on the lean adjoint ODE (contribution (ii)) are load-bearing for the theoretical support of adjoint methods; the proof should clarify whether the bounds remain uniform in the dimension of the data space, as high-dimensional image models (e.g., Stable Diffusion) could otherwise render the constants prohibitive.

    Authors: The derived norm bounds on the lean adjoint ODE are uniform in the data dimension. The proof proceeds via a Gronwall inequality applied to the adjoint dynamics whose growth is controlled by the Lipschitz constant of the vector field; this constant is independent of ambient dimension under the standard smoothness assumptions on the score/flow. Consequently the final bound is parameter-free and dimension-independent, which is why it remains informative for high-dimensional models such as Stable Diffusion. We will add an explicit remark in Section 5 and in the proof appendix stating the dimension uniformity and briefly discussing its relevance to image-scale applications. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper unifies existing SOC and non-equilibrium thermodynamics approaches to exponential tilting for diffusion/flow models, then derives independent contributions: explicit bias-variance decompositions for gradient estimators, norm bounds on the lean adjoint ODE, and adaptations of CMCD/NETS losses plus new Crooks/Jarzynski identities. These are presented as fresh analytic results rather than reductions of prior fitted quantities or self-citations. Experimental validation on Stable Diffusion 1.5/3 provides external falsifiability. No quoted step equates a claimed prediction or theorem to its own inputs by construction, and no load-bearing premise collapses to a self-citation chain. The framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available so the ledger is necessarily incomplete; the work extends existing SOC and thermodynamics frameworks without apparent new free parameters or invented entities visible at this level.

axioms (1)
  • standard math Standard assumptions of stochastic optimal control and non-equilibrium thermodynamics for diffusion and flow processes hold in the tilting setting.
    The unified framework is built directly on these established mathematical perspectives.

pith-pipeline@v0.9.0 · 5478 in / 1358 out tokens · 40303 ms · 2026-05-09T19:40:32.037777+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

300 extracted references · 28 canonical work pages · 4 internal anchors

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    DiffusionNFT: Online Diffusion Reinforcement with Forward Process

    Diffusionnft: Online diffusion reinforcement with forward process , author=. arXiv preprint arXiv:2509.16117 , year=

    work page internal anchor Pith review arXiv

  3. [3]

    Physical Review Letters , volume=

    Escorted free energy simulations: Improving convergence by reducing dissipation , author=. Physical Review Letters , volume=. 2008 , publisher=

    2008

  4. [4]

    Fine-tuning of continuous- time diffusion models as entropy-regularized control.arXiv preprint arXiv:2402.15194,

    Fine-tuning of continuous-time diffusion models as entropy-regularized control , author=. arXiv preprint arXiv:2402.15194 , year=

    work page arXiv

  5. [5]

    Physical Review E , volume=

    Time-asymmetric fluctuation theorem and efficient free-energy estimation , author=. Physical Review E , volume=. 2024 , publisher=

    2024

  6. [6]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  7. [7]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Fast two-sample testing with analytic representations of probability measures , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    International Conference on Machine Learning , pages=

    Particle Denoising Diffusion Sampler , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  10. [10]

    The Twelfth International Conference on Learning Representations , year=

    Reverse Diffusion Monte Carlo , author=. The Twelfth International Conference on Learning Representations , year=

  11. [11]

    Huang, Jian and Jiao, Yuling and Kang, Lican and Liao, Xu and Liu, Jin and Liu, Yanyan , journal=. Schr

  12. [12]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Iterated denoising energy matching for sampling from boltzmann densities , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  13. [13]

    arXiv:1707.07269 [math]

    Large sample analysis of the median heuristic , author=. arXiv preprint arXiv:1707.07269 , year=

    work page arXiv

  14. [14]

    International Conference on Artificial Intelligence and Statistics , pages=

    Two-sample testing using deep learning , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2020 , organization=

    2020

  15. [15]

    International Conference on Learning Representations , year=

    Learning Neural Event Functions for Ordinary Differential Equations , author=. International Conference on Learning Representations , year=

  16. [16]

    Journal of Machine Learning Research , volume=

    Monte carlo gradient estimation in machine learning , author=. Journal of Machine Learning Research , volume=

  17. [17]

    International Conference on Machine Learning , pages=

    Do differentiable simulators give better policy gradients? , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  18. [18]

    Proceedings of the International Conference on Machine Learning , year=

    The Mechanics of n -Player Differentiable Games , author=. Proceedings of the International Conference on Machine Learning , year=

  19. [19]

    Elementary Principles in Statistical Mechanics: Developed with Especial Reference to the Rational Foundation of Thermodynamics , publisher=

    Gibbs, Josiah Willard , year=. Elementary Principles in Statistical Mechanics: Developed with Especial Reference to the Rational Foundation of Thermodynamics , publisher=

  20. [20]

    A tutorial on energy-based learning , author=

  21. [21]

    Diffusion Models Beat GANs on Image Synthesis

    Diffusion Models Beat GANs on Image Synthesis , author=. arXiv preprint arXiv:2105.05233 , year=

    work page internal anchor Pith review arXiv

  22. [22]

    Adversarial score matching and improved sampling for image generation

    Adversarial score matching and improved sampling for image generation , author=. arXiv preprint arXiv:2009.05475 , year=

    work page arXiv 2009

  23. [23]

    Solving linear inverse problems using the prior implicit in a denoiser

    Solving linear inverse problems using the prior implicit in a denoiser , author=. arXiv preprint arXiv:2007.13640 , year=

    work page arXiv 2007

  24. [24]

    Porretta, Alessio , year =. Weak. doi:10/f64gfr , journal =

  25. [25]

    Auto-Encoding Variational Bayes

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Efficient learning of sparse representations with an energy-based model , author=

  27. [27]

    Generative modeling with denoising auto- encoders and langevin sampling.arXiv preprint arXiv:2002.00107,

    Generative modeling with denoising auto-encoders and Langevin sampling , author=. arXiv preprint arXiv:2002.00107 , year=

    work page arXiv 2002

  28. [28]

    Conference on Learning Theory , pages=

    Implicit bias of gradient descent for wide two-layer neural networks trained with the logistic loss , author=. Conference on Learning Theory , pages=. 2020 , organization=

    2020

  29. [29]

    Conference on Learning Theory , year=

    How do infinite width bounded norm networks look in function space? , author=. Conference on Learning Theory , year=

  30. [30]

    International Conference on Learning Representations (ICLR 2020) , year=

    A Function Space View of Bounded Norm Infinite Width ReLU Nets: The Multivariate Case , author=. International Conference on Learning Representations (ICLR 2020) , year=

    2020

  31. [31]

    Conference on Learning Theory , year=

    Kernel and rich regimes in overparametrized models , author=. Conference on Learning Theory , year=

  32. [32]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Implicit generation and generalization in energy-based models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  33. [33]

    Lei, Qi and Lee, Jason D and Dimakis, Alexandros G and Daskalakis, Constantinos , journal=

  34. [34]

    arXiv preprint arXiv:1812.02878 , year=

    Solving non-convex non-concave min-max games under PolyakLojasiewicz condition , author=. arXiv preprint arXiv:1812.02878 , year=

    work page arXiv

  35. [35]

    Unbalanced Optimal Transport: Dynamic and

    Chizat, Lenaic and Peyré, Gabriel and Schmitzer, Bernhard and Vialard, François-Xavier , journal=. Unbalanced Optimal Transport: Dynamic and

  36. [36]

    2018 , author =

    Unbalanced optimal transport: Dynamic and Kantorovich formulations , journal =. 2018 , author =

    2018

  37. [37]

    A new optimal transport distance on the space of finite

    Kondratyev, Stanislav and Monsaingeon, L. A new optimal transport distance on the space of finite. Advances in Differential Equations , volume=. 2016 , publisher=

    2016

  38. [38]

    1967 , month = jan, volume =

    Local Behavior of Solutions of Quasilinear Parabolic Equations , author =. 1967 , month = jan, volume =. doi:10/dwv9tq , journal =

    1967

  39. [39]

    International Conference on Learning Representations , year=

    Stable Opponent Shaping in Differentiable Games , author=. International Conference on Learning Representations , year=

  40. [40]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Multi-Agent Learning with Policy Prediction , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  41. [41]

    Wasserstein GAN

    Wasserstein gan , author=. arXiv preprint arXiv:1701.07875 , year=

    work page Pith review arXiv

  42. [42]

    Inverse problems , volume=

    Stable architectures for deep neural networks , author=. Inverse problems , volume=. 2017 , publisher=

    2017

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    Reinforcement learning for fine-tuning text-to-image diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Advances in neural information processing systems , volume=

    Linearly-solvable Markov decision problems , author=. Advances in neural information processing systems , volume=

  45. [45]

    Proceedings of the International Conference on Autonomous Agents and Multi-Agent Systems , year=

    Learning with Opponent-Learning Awareness , author=. Proceedings of the International Conference on Autonomous Agents and Multi-Agent Systems , year=

  46. [46]

    Make-A-Video: Text-to-Video Generation without Text-Video Data

    Make-a-video: Text-to-video generation without text-video data , author=. arXiv preprint arXiv:2209.14792 , year=

    work page internal anchor Pith review arXiv

  47. [47]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [48]

    Forty-first International Conference on Machine Learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first International Conference on Machine Learning , year=

  49. [49]

    Advances in neural information processing systems , volume=

    Voicebox: Text-guided multilingual universal speech generation at scale , author=. Advances in neural information processing systems , volume=

  50. [50]

    Audiobox: Unified audio generation with natural language prompts.arXiv preprint arXiv:2312.15821,

    Audiobox: Unified audio generation with natural language prompts , author=. arXiv preprint arXiv:2312.15821 , year=

    work page arXiv

  51. [51]

    Foundations and Trends in Machine Learning , volume =

    Bubeck, Sébastien , title =. Foundations and Trends in Machine Learning , volume =

  52. [52]

    Entropy , volume=

    Interacting particle solutions of fokker--planck equations through gradient--log--density estimation , author=. Entropy , volume=. 2020 , publisher=

    2020

  53. [53]

    Stochastic Systems , volume =

    Juditsky, Anatoli and Nemirovski, Arkadi and Tauvel, Claire , title =. Stochastic Systems , volume =

  54. [54]

    4or , volume=

    Generalized Nash equilibrium problems , author=. 4or , volume=

  55. [55]

    Mathematical Programming , volume=

    Finite-dimensional variational inequality and nonlinear complementarity problems: a survey of theory, algorithms and applications , author=. Mathematical Programming , volume=. 1990 , publisher=

    1990

  56. [56]

    Matecon , volume=

    The extragradient method for finding saddle points and other problems , author=. Matecon , volume=

  57. [57]

    Prox-method with rate of convergence O(1/t) for variational inequalities with

    Nemirovski, Arkadi , journal=. Prox-method with rate of convergence O(1/t) for variational inequalities with. 2004 , publisher=

    2004

  58. [58]

    Mathematical Programming , volume=

    Dual extrapolation and its applications to solving variational inequalities and related problems , author=. Mathematical Programming , volume=. 2007 , publisher=

    2007

  59. [59]

    53rd IEEE Conference on Decision and Control , pages=

    Optimal robust smoothing extragradient algorithms for stochastic variational inequality problems , author=. 53rd IEEE Conference on Decision and Control , pages=. 2014 , organization=

    2014

  60. [60]

    SIAM Journal on Optimization , volume=

    Extragradient method with variance reduction for stochastic variational inequalities , author=. SIAM Journal on Optimization , volume=. 2017 , publisher=

    2017

  61. [61]

    International Conference on Learning Representations , year=

    Optimistic mirror descent in saddle-point problems: Going the extra (gradient) mile , author=. International Conference on Learning Representations , year=

  62. [62]

    Set-Valued and Variational Analysis , volume=

    On stochastic mirror-prox algorithms for stochastic Cartesian variational inequalities: Randomized block coordinate and optimal averaging schemes , author=. Set-Valued and Variational Analysis , volume=. 2018 , publisher=

    2018

  63. [63]

    arXiv preprint arXiv:1810.10207 , year=

    Solving weakly-convex-weakly-concave saddle-point problems as weakly-monotone variational inequality , author=. arXiv preprint arXiv:1810.10207 , year=

    work page arXiv

  64. [64]

    Advances in Neural Information Processing Systems , pages=

    Solving a class of non-convex min-max games using iterative first order methods , author=. Advances in Neural Information Processing Systems , pages=

  65. [65]

    Advances in Neural Information Processing Systems , pages=

    The limit points of (optimistic) gradient descent in min-max optimization , author=. Advances in Neural Information Processing Systems , pages=

  66. [66]

    arXiv preprint arXiv:1805.05751 , year=

    Local saddle point optimization: A curvature exploitation approach , author=. arXiv preprint arXiv:1805.05751 , year=

    work page arXiv

  67. [67]

    Reducing Noise in

    Chavdarova, Tatjana and Gidel, Gauthier and Fleuret, Francois and Foo, Chuan-Sheng and Lacoste-Julien, Simon , booktitle=. Reducing Noise in

  68. [68]

    Advances in Neural Information Processing Systems , pages=

    Stochastic variance reduction methods for saddle-point problems , author=. Advances in Neural Information Processing Systems , pages=

  69. [69]

    Proceedings of the Conference on Learning Theory , year=

    A universal algorithm for variational inequalities adaptive to smoothness and noise , author=. Proceedings of the Conference on Learning Theory , year=

  70. [70]

    2001 , booktitle=

    Rational and convergent learning in stochastic games , author=. 2001 , booktitle=

    2001

  71. [71]

    2007 , publisher=

    Conitzer, Vincent and Sandholm, Tuomas , journal=. 2007 , publisher=

    2007

  72. [72]

    Training

    Daskalakis, Constantinos and Ilyas, Andrew and Syrgkanis, Vasilis and Zeng, Haoyang , booktitle=. Training

  73. [73]

    The numerics of

    Mescheder, Lars and Nowozin, Sebastian and Geiger, Andreas , booktitle=. The numerics of

  74. [74]

    Gradient descent

    Nagarajan, Vaishnavh and Kolter, J Zico , booktitle=. Gradient descent

  75. [75]

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle=

  76. [76]

    International Conference on Learning Representations , year=

    A variational inequality perspective on generative adversarial networks , author=. International Conference on Learning Representations , year=

  77. [77]

    On finding local

    Mazumdar, Eric V and Jordan, Michael I and Sastry, S Shankar , journal=. On finding local

  78. [78]

    Advances in Neural Information Processing Systems , pages=

    Generative adversarial nets , author=. Advances in Neural Information Processing Systems , pages=

  79. [79]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops , pages=

    Curiosity-driven exploration by self-supervised prediction , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops , pages=

  80. [80]

    Advances in Neural Information Processing Systems , pages=

    Imagination-augmented agents for deep reinforcement learning , author=. Advances in Neural Information Processing Systems , pages=

Showing first 80 references.