Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Calibrating Generative Models to Distributional Constraints

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Calibrating a generative model to distributional constraints is a search for the closest model in KL divergence, and two surrogate losses make that search practical at scale.

desk verdict A genuinely useful constrained fine-tuning method with solid math and strong large-scale demos, but the abstract overstates the model size and the gradient estimator's variance is under-examined. read the letter →

arxiv 2510.10020 v4 pith:3CLMCF7F submitted 2025-10-11 stat.ML cs.LGq-bio.BM

Henry D. Smith , Nathaniel L. Diamant , Brian L. Trippe This is my paper
classification stat.MLcs.LGq-bio.BM
keywords calibrationgenerativemodelsdistributionalconstraintsKLdivergencemaximumentropyfine-tuningscorefunctiongradientsproteindesign
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to show that calibration of a generative model — making the statistics of its samples match specified values, such as class proportions, gender balance, or protein secondary-structure content — is a constrained optimization problem: find the model closest to the base distribution in KL divergence whose expectation of a chosen statistic equals the target. Because exact constraints are intractable, the authors propose two surrogate fine-tuning objectives. The relax loss penalizes constraint violation while adding a KL penalty; the reward loss estimates the maximum-entropy solution, the exponential tilt p ∝ p_base exp(α^T h(x)), and fine-tunes the model toward that target. They report that these objectives substantially reduce calibration error under hundreds of simultaneous constraints on models with more than a billion parameters across protein design, image generation, and language modeling, and prove that the relax-loss optimum converges to the maximum-entropy solution as the penalty weight shrinks.

What carries the argument

The engine is the maximum-entropy exponential-tilt identity, p_α*(x) ∝ p_base(x) exp(α^T h(x)), which characterizes the KL-closest distribution satisfying the moment constraint. The reward loss is built by estimating α* through the concave dual problem α_hat = argmax_α α^T h* − log(mean_n exp(α^T h(x_n))) and then minimizing D_KL(p_θ ∥ p_α_hat) — the same loss used in reward fine-tuning, with r(x)=α_hat^T h(x). The relax loss instead minimizes ||E_pθ[h] − h*||² + λ D_KL(p_θ ∥ p_base), with a bias-corrected Monte Carlo estimate of the squared violation. All gradients are score-function estimators: they differentiate the importance weight p_θ(x)/p_θ′(x) evaluated at a detached copy of the para

What would settle it

Replicate the 100-dimensional GMM calibration experiment from Section 3 with the paper's batch size (10^4) and optimization schedule, and measure the constraint violation after 2×10^3 steps; if the majority of the initial constraint violation is not eliminated, the claim that CGM-relax scales to hundreds of constraints fails.

Watch

Extended reading notes

Core claim

The central discovery is a reduction: imposing distribution-level constraints on a generative model is equivalent to a KL-constrained search, and its solution has a closed-form exponential-tilt representation. For a base model p_base and statistic h, the unique distribution closest to p_base in KL divergence subject to E_p[h(x)] = h* is p_α*(x) ∝ p_base(x) exp(α*^T h(x)). The paper shows how to estimate α* from a sample of the base model by solving a concave dual problem, and then converts the calibration task into two tractable surrogate losses: a penalty loss on the squared deviation of the expected statistic, and a reward loss that is exactly the objective of reward fine-tuning. The paper

Load-bearing premise

The whole method rests on the score-function gradient estimators being unbiased and low-variance; the paper proves unbiasedness under a square-integrability condition but never measures the estimator's variance on real models, and its own toy experiments show instability for rare events at small batch sizes.

Editorial extensions

If this is right

  • Practitioners can fine-tune a single pretrained model to satisfy many expectation constraints at once — proportional class outputs, fairness targets, structural diversity — without designing a bespoke loss for each domain.
  • The relax loss scales to hundreds or a thousand simultaneous constraints (the paper demonstrates up to 10^3 in a simulation); the reward loss is limited by the feasibility of the empirical max-entropy dual, which the paper finds fails beyond roughly 30 constraints in its simulations.
  • The regularization parameter λ gives an explicit trade-off: lower λ yields tighter constraint satisfaction at the cost of larger KL deviation from the base model, letting users choose the operating point.
  • Because the reward loss coincides with reward fine-tuning, the calibration problem inherits the available machinery for preference optimization, broadening the tools that can be brought to bear.
  • The approach requires tractable likelihoods of both the model and the base; it does not directly apply to implicit generative models such as GANs or VAEs, a boundary the paper states explicitly.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The failure of the reward route beyond ~30 constraints implies that estimating α* by importance weighting a finite base-model sample inherits the curse of dimensionality; a learned density-ratio or energy-based estimator for the tilt could extend the reward approach to far more constraints.
  • The rare-event degradation in the toy GMM (below base probability ≈10⁻³ at batch size 100) points to gradient-estimator variance as the binding constraint, not the loss formulation; larger batches, control variates, or pathwise estimators could push calibration into the rare-tail regime.
  • The TarFlow FID increase after calibration suggests that expectation constraints alone do not preserve sample realism; adding a quality penalty or enforcing constraints in a feature space that already encodes realism would be a natural, testable extension.
  • If the scaling results hold, calibration becomes a post-hoc safety tool: rather than retraining, one fine-tunes a frozen pretrained model with a handful of expectation constraints, which may be far more data-efficient than full alignment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper proposes two fine-tuning methods, CGM-relax and CGM-reward, for calibrating a pretrained generative model to expectation constraints. The formal target is the KL projection of the base model onto the set of distributions satisfying E_{p_θ}[h]=h*. CGM-relax replaces the constraint by a quadratic miscalibration penalty plus a λ-weighted KL term; CGM-reward estimates the parameters α of the exponential-tilt maximum-entropy solution and then minimizes reverse KL to that target. The authors derive unbiased loss and score-function gradient estimators, prove consistency and asymptotic normality of the α estimator, and evaluate the methods on a diffusion GMM toy problem, two protein models (Genie2, ESM3-open), a normalizing-flow image model (TarFlow), and TinyStories-33M. The paper reports substantial calibration-error reductions in several settings, while also acknowledging residual error and the restriction to models with tractable likelihoods.

Significance. The core theory is a useful contribution: Appendix B gives careful unbiasedness arguments for the variance-corrected L_viol estimator and the importance-weighted score gradients, and Appendix C connects both losses to maximum entropy, with consistency and rate results for the α estimator. The synthetic experiments usefully isolate the roles of λ, N, and constraint dimension. If the empirical results withstand scrutiny, the paper offers a practical, relatively model-agnostic recipe for likelihood-based generative models. However, the paper’s strongest advertised claims—that “these approaches” handle hundreds of simultaneous constraints and models up to 9B/1B parameters—are not supported by the evidence: CGM-reward fails beyond roughly 30 constraints in the toy setting, and the largest model actually evaluated is 1.4B. The lack of gradient-variance diagnostics also leaves open whether the reported gains are stable. The released code and the careful appendices are notable strengths and increase the value of the work.

major comments (3)
  1. [Abstract; Section 4, Table 1; Section 3, Fig. 2B] The arXiv abstract claims calibration “across hundreds of simultaneous constraints and models with up to nine billion parameters,” while the main-text abstract and conclusion say “up to one billion.” The largest model actually evaluated is ESM3-open with 1.4B parameters (Table 1). Moreover, the “hundreds of constraints” result is specific to CGM-relax: the toy experiment in Fig. 2B shows that CGM-reward is infeasible for more than about 30 constraints, and the case studies use 99 (CGM-relax) or 15 (CGM-reward) protein constraints, 5 TarFlow constraints, and 8 TinyStories constraints. Please revise the abstract and conclusion to state which method scales and to give the true parameter counts.
  2. [Section 3, Fig. 2B; Section 2.3; Appendix B.2] The paper does not provide variance diagnostics for the score-function gradient estimators in the real-model runs, and the toy evidence suggests this matters. In Fig. 2B, CGM-reward is outperformed by CGM-relax at high constraint dimension even when α* is fixed to an oracle value, so the failure cannot be attributed solely to the dual infeasibility discussed in Appendix C.2. In the product-GMM toy, the exponential tilt of the base model is itself a product GMM, so the reverse-KL surrogate is reasonably well specified; the remaining gap therefore points to optimization or gradient-variance issues. No seed-to-seed variance is reported for TarFlow (Section 4.2), and Genie2 results are averages over only 3 trials. Please add gradient-noise and run-to-run variability diagnostics for the main experiments, or weaken the claim that CGM-reward reliably scales to high-dimensional settings.
  3. [Introduction (p. 1); Appendix E.4.3] The introduction states that neither Khalifa et al. (2021) nor Shen et al. (2024) “reduces a majority of calibration error,” but Appendix E.4.3 reports that the Khalifa et al. baseline reduces miscalibration by roughly 85% in the TinyStories experiment. This is an internal inconsistency. If the claim is meant to apply only to the present multi-application evaluation rather than to the cited methods in general, it should be qualified accordingly.
minor comments (5)
  1. [Section 2.3, Eq. (10)] The notation ∇_θ[p_θ(x_m)f(x_m,θ)] should be expanded explicitly, since f may depend on θ; the current display is terse. In Algorithms 1 and 2, l_m is defined with stop-grad(θ), while in Appendix B.2 the same symbol uses θ; aligning these notations would help readers.
  2. [Appendix C.3, Proposition C.6] The statement says “lim_{λ→∞}”, but the proof and the surrounding text use λ→0. The limit should be λ→0; as written the theorem contradicts the finite-λ bound that follows it.
  3. [Table 1] The row for xSpace is difficult to read due to formatting (e.g., “100 (S^{100}_{4096}) 50”). Please clean up the table formatting.
  4. [Section 2.2] The statement that the parametric optimum p_θ* will be close to p_α* “if the class of generative models is sufficiently expressive” is informal. A short remark clarifying that no quantitative bound is provided would be helpful, since the theory in Appendix C is nonparametric.
  5. [Section 3, Fig. 2A] The text says performance degrades below π=1e-3, but the figure/axis labels are not fully described in the caption. Consider adding precise evaluation metrics and error bars for the rare-event sweep.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: relax and reward objectives are independently characterized surrogates, and self-citations are contextual only.

full rationale

The central derivation chain is self-contained and not circular. The calibration problem (Eq. 1) is a well-defined constrained optimization; CGM-relax (Eq. 2) is an explicit penalty surrogate with the miscalibration term Lviol plus λ·D_KL(pθ∥pθbase), and CGM-reward is constructed from the maximum-entropy tilt pα* ∝ pθbase exp(α⊤h), with α estimated by the empirical dual problem (Eq. 7). The key theoretical claims—the form and uniqueness of the maximum-entropy solution (Theorem 2.1 / C.3), the dual characterization (Eq. 6), the consistency of bαN (C.4), and the connection of the relax loss to a tilted distribution with rate O(λ) (C.3)—are proven in Appendices B–C using standard convex duality and external results, not by assuming the paper's own conclusions. The reduction in miscalibration is the optimization objective itself, not a hidden fitted parameter relabeled as a prediction; λ and N are disclosed hyperparameters whose tradeoffs are explicitly studied in Section 3. The rare-event degradation and the restriction to tractable-likelihood models are acknowledged limitations (Section 3 and Section 5), and the lack of variance diagnostics is an empirical robustness concern, not evidence of circularity. Self-citations to the authors' prior work (e.g., Trippe et al. 2023, Watson et al. 2023, Uehara et al. 2024) appear only as application context or related-work background, and they are not load-bearing for the paper's central derivation or its main empirical claims.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

Central method rests on unverified expressiveness and gradient-variance assumptions; lambda, N, and M are hand-set hyperparameters that directly control the reported results.

free parameters (3)
  • lambda (relax regularization) = 1e-3 (Genie2), 1e-2 (ESM3), 1e-4 (TarFlow), 0.1 (TinyStories)
    Grid-searched per application to balance miscalibration penalty and KL. The relax solution and all reported reductions depend on this choice; theory only gives lambda -> 0 behavior.
  • N (base samples for alpha_hat_N) = 1e5 (simulations), 2.5e4 (Genie2 per Appendix E.1; main text says 2.5e3), 5e3 (TarFlow)
    Hand-chosen; CGM-reward's feasibility and variance depend on N, and infeasibility grows with constraint dimension.
  • M (batch size) = 1e4 (simulations), 64-512 in case studies
    Gradient estimator variance scales with batch size; rare-event performance collapses at M=100 for pi <= 1e-3 (Section 3).
assumptions (6)
  • standard math Interior moment condition (Assumption C.1)
    Required for existence of a max-entropy solution of exponential-tilt form; used in Theorem 2.1 and Appendix C.
  • standard math Open domain of log-normalizer (Assumption C.2)
    Regular exponential-family assumption needed for differentiability and duality.
  • standard math No linear redundancy in h (Assumption C.4)
    Ensures a unique dual optimum; used for consistency and asymptotic normality of alpha_hat_N.
  • domain assumption Model class is expressive enough to approximate the nonparametric max-entropy solution
    Section 2.2 and Appendix C.3 assume the parametric family p_theta can approximate p_alpha*; not verified for Genie2/ESM3/TarFlow.
  • domain assumption Gradient estimator variance is manageable
    Unbiasedness (Propositions B.4/B.5) does not guarantee finite variance; Section 3 shows degradation for rare events.
  • domain assumption Tractable likelihoods and densities with respect to the base model
    Algorithms require sampling from p_theta and computing log p_theta and its gradients; excludes VAEs/GANs, acknowledged in the conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Calibrating Generative Models to Distributional Constraints." pith.science (2026). https://pith.science/paper/3CLMCF7F

@misc{pith2026251010020,
  author       = {Pith},
  title        = {Pith review of: Calibrating Generative Models to Distributional Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CLMCF7F}},
  note         = {Machine review of arXiv:2510.10020}
}
read the original abstract

Generative models frequently suffer miscalibration, wherein statistics of the sampling distribution, such as the fraction of generations in a given class, deviate from desired values. We frame calibration as a constrained optimization problem and seek the closest model in Kullback-Leibler divergence satisfying a calibration constraint. To address the intractability of imposing these constraints exactly, we introduce two surrogate objectives for fine-tuning: (1) the relax loss, which replaces the constraint with a miscalibration penalty, and (2) the reward loss, which converts calibration into a reward fine-tuning problem. We demonstrate that these approaches substantially reduce calibration error across hundreds of simultaneous constraints and models with up to nine billion parameters, spanning applications in protein design, image generation, and language modeling.

Figures

Figures reproduced from arXiv: 2510.10020 by the authors.

Figure 1
Figure 1. Calibrating mixture proportions in a 1D GMM. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. A: CGM effectively upweights the probability of a rare mode in a 1D GMM. B: CGM￾relax calibrates the base model to up to 103 constraints, whereas CGM-reward is not well-defined for >30 constraints. When αbN is fixed to α∗ (red dashed line), CGM-relax outperforms CGM-reward. the calibration problem to high accuracy. In the remaining experiments, we perform grid-search to select λ in CGM-relax and use N = 105 samples … view at source ↗
Figure 3
Figure 3. A: Samples from the Genie2 protein generative models before and after calibration with CGM-relax (λ=10−3 ). B: CGM-relax reduces the distance of secondary structure content to natural proteins by >4 times for Genie2 and >2 times for ESM3 while maintaining biophysical plausibility. Across all examples, CGM reduces the majority of calibration error without significantly degrading the quality of generations. Consistent… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Generations from the conditional TarFlow model (Zhai et al., 2025) before and after [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: A: Gender imbalance and distance from base-model (symmetrized KL from pre-trained TinyStories-33M). B: Gender imbalance for professions included and heldout from calibration be￾fore and after CGM-relax (λ = 0.1). Points below the diagonal were improved by CGM. 4.3 ELIM…
Figure 6
Figure 6. Figure 6: Random samples from the Genie2 model before calibration (top), after calibration using [PITH_FULL_IMAGE:figures/full_fig_p035_6.png]
Figure 7
Figure 7. Figure 7: Random samples from the ESM3-open model before calibration (top), after calibration [PITH_FULL_IMAGE:figures/full_fig_p038_7.png]
Figure 8
Figure 8. Figure 8: Random samples from the conditional TarFlow model trained on the AFHQ dataset (blue [PITH_FULL_IMAGE:figures/full_fig_p039_8.png]
Figure 9
Figure 9. Figure 9: Calibrating TarFlow with CGM-relax reduces the TV distance of animal class labels to [PITH_FULL_IMAGE:figures/full_fig_p040_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    DRL trains a discriminator on data versus base-model samples in pretrained representation space and uses its logit as reward in KL-regularized RL, cutting guidance-free FID from 9.38 to 2.62 on SiT and similar gains o...

  2. Na\"ive PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation

    cs.CV 2026-03 conditional novelty 6.0 of 10

    A lightweight predictor ranks initial noises by expected human-preference score for a prompt, selecting the best few for diffusion generation and reporting prompt difficulty.

  3. Can We Predict The Human Preference For Text-to-Image Content Prior To Generation And Is It Even Useful To Do So?

    cs.CV 2026-06 unverdicted novelty 3.0 of 10

    Exploration of pre-generation prediction of human preference metrics (HPM) from noise seeds in diffusion models to improve output quality with negligible added cost.

Reference graph

Works this paper leans on

82 extracted references · 6 linked inside Pith · cited by 3 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    ABA profile of the legal profession 2024 — demographics, 2024

    American Bar Association . ABA profile of the legal profession 2024 — demographics, 2024

    2024

  3. [3]

    Reverse-time diffusion equation models

    Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 1982

    1982

  4. [5]

    Lecture notes in convex analysis, nonlinear programming theory, and nonlinear programming algorithms, 2023

    Aharon Ben-Tal and Arkadi Nemirovski. Lecture notes in convex analysis, nonlinear programming theory, and nonlinear programming algorithms, 2023

    2023

  5. [6]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In International Conference on Learning Representations, 2024

    2024

  6. [7]

    Techniques of variational analysis

    Jonathan M Borwein and Qiji J Zhu. Techniques of variational analysis. Springer, 2005

    2005

  7. [8]

    Integrating molecular simulation and experimental data: a B ayesian/maximum entropy reweighting approach

    Sandro Bottaro, Tone Bengtsen, and Kresten Lindorff-Larsen. Integrating molecular simulation and experimental data: a B ayesian/maximum entropy reweighting approach. Structural Bioinformatics: Methods and Protocols, 2020

    2020

  8. [9]

    Convex optimization

    Stephen P Boyd and Lieven Vandenberghe. Convex optimization. Cambridge University Press, 2004

    2004

Show all 82 references
  1. [10]

    Transformations of W iener integrals under translations

    Robert H Cameron and William T Martin. Transformations of W iener integrals under translations. Annals of Mathematics, 1944

    1944

  2. [11]

    Constrained discrete diffusion

    Michael Cardei, Jacob K Christopher, Thomas Hartvigsen, Brian R Bartoldson, Bhavya Kailkhura, and Ferdinando Fioretto. Constrained discrete diffusion. arXiv preprint arXiv:2503.09790, 2025

    2025

  3. [12]

    StarGAN v2: Diverse image synthesis for multiple domains

    Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020

    2020

  4. [13]

    Deep reinforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in Neural Information Processing Systems, 2017

    2017

  5. [14]

    I-divergence geometry of probability distributions and minimization problems

    Imre Csisz \'a r. I-divergence geometry of probability distributions and minimization problems. The Annals of Probability, 1975

    1975

  6. [15]

    Ragotte, Lukas F

    Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J. Ragotte, Lukas F. Milles, Basile I. M. Wicky, Alexis Courbet, Rob J. de Haas, Neville Bethel, Philip J. Y. Leung, Timothy F. Huddy, Sam Pellock, Doug Tischer, F. Chan, Brian Koepnick, H. Nguyen, A. Kang, ...

    2022

  7. [16]

    The well-calibrated B ayesian

    A Philip Dawid. The well-calibrated B ayesian. Journal of the American Statistical Association, 1982

    1982

  8. [17]

    DEFT : efficient fine-tuning of diffusion models by learning the generalised h -transform

    Alexander Denker, Francisco Vargas, Shreyas Padhy, Kieran Didi, Simon Mathis, Riccardo Barbano, Vincent Dutordoir, Emile Mathieu, Urszula Julia Komorowska, and Pietro Lio. DEFT : efficient fine-tuning of diffusion models by learning the generalised h -transform. Advances in Ne...

    2024

  9. [18]

    Diffusion models beat GAN s on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GAN s on image synthesis. Advances in Neural Information Processing Systems, 2021

    2021

  10. [19]

    Adjoint matching: fine-tuning flow and diffusion generative models with memoryless stochastic optimal control

    Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky TQ Chen. Adjoint matching: fine-tuning flow and diffusion generative models with memoryless stochastic optimal control. In International Conference on Learning Representations, 2025

    2025

  11. [20]

    An image is worth 16x16 words: transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: transformers for image recognition at scale. In International Con...

    2021

  12. [21]

    T iny S tories: How small can language models be and still speak coherent E nglish? arXiv preprint arXiv:2305.07759, 2023

    Ronen Eldan and Yuanzhi Li. T iny S tories: How small can language models be and still speak coherent E nglish? arXiv preprint arXiv:2305.07759, 2023

    2023 arXiv

  13. [22]

    DPOK : Reinforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. DPOK : Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 2023

    2023

  14. [23]

    F air D iffusion: Instructing text-to-image generation models on fairness

    Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. F air D iffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023

    2023 arXiv

  15. [24]

    Bias and fairness in large language models: A survey

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 2024

    2024

  16. [25]

    On transforming a certain class of stochastic processes by absolutely continuous substitution of measures

    Igor Vladimirovich Girsanov. On transforming a certain class of stochastic processes by absolutely continuous substitution of measures. Theory of Probability & Its Applications, 1960

    1960

  17. [26]

    Aligning language models with preferences through f-divergence minimization

    Dongyoung Go, Tomasz Korbak, Germ \'a n Kruszewski, Jos Rozen, Nahyeon Ryu, and Marc Dymetman. Aligning language models with preferences through f-divergence minimization. In International Conference on Machine Learning, 2023

    2023

  18. [27]

    Generative AI and Biology (GenBio)

    Sven Gutjahr, Riccardo De Santi, Luca Schaufelberger, Kjell Jorner, and Andreas Krause. Constrained molecular generation via sequential flow model fine-tuning. In “Generative AI and Biology (GenBio)” Workshop the International Conference on Machine Learning, 2025

    2025

  19. [28]

    Simulating 500 million years of evolution with a language model

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q Tran, Jonathan Deaton, Marius Wiggert, Rohil Badkundri, Irhum Shafkat, Jun Gong, Alexander Derry, Raul S Molina, Neil Thomas, Yousuf A Khan, Chetan Mishra, Carolyn Ki...

    2025

  20. [29]

    Deep Generative Models and Downstream Applications

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. “Deep Generative Models and Downstream Applications” Workshop at the Advances in Neural Information Processing Systems Conference, 2021

    2021

  21. [30]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020

    2020

  22. [31]

    Generative models for graph-based protein design

    John Ingraham, Vikas Garg, Regina Barzilay, and Tommi Jaakkola. Generative models for graph-based protein design. Advances in Neural Information Processing Systems, 2019

    2019

  23. [32]

    Information theory and statistical mechanics

    Edwin T Jaynes. Information theory and statistical mechanics. Physical Review, 1957

    1957

  24. [33]

    Constrained diffusion models via dual training

    Shervin Khalafi, Dongsheng Ding, and Alejandro Ribeiro. Constrained diffusion models via dual training. Advances in Neural Information Processing Systems, 2024

    2024

  25. [34]

    A distributional approach to controlled text generation

    Muhammad Khalifa, Hady Elsahar, and Marc Dymetman. A distributional approach to controlled text generation. In International Conference on Learning Representations, 2021

    2021

  26. [35]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015

    2015

  27. [36]

    An information-theoretic alternative to generalized method of moments estimation

    Yuichi Kitamura and Michael Stutzer. An information-theoretic alternative to generalized method of moments estimation. Econometrica: Journal of the Econometric Society, 1997

    1997

  28. [37]

    u rgen K \

    J \"u rgen K \"o finger, Lukas S. Stelzl, Klaus Reuter, C \'e sar Allande, Katrin Reichel, and Gerhard Hummer. Efficient ensemble refinement by reweighting. Journal of Chemical Theory and Computation, 2019

    2019

  29. [38]

    Deep RL Meets Structured Prediction

    Wouter Kool, Herke van Hoof, and Max Welling. Buy 4 REINFORCE samples, get a baseline for free! In “Deep RL Meets Structured Prediction” Workshop at the International Conference on Learning Representations , 2019

    2019

  30. [39]

    Information Theory and Statistics

    Solomon Kullback. Information Theory and Statistics. John Wiley & Sons, 1959

    1959

  31. [40]

    B iotite: a unifying open source computational biology framework in P ython

    Patrick Kunzmann and Kay Hamacher. B iotite: a unifying open source computational biology framework in P ython. BMC Bioinformatics, 2018

    2018

  32. [41]

    A perspective on the use of control variables to increase the efficiency of M onte C arlo simulations

    Stephen S Lavenberg and Peter D Welch. A perspective on the use of control variables to increase the efficiency of M onte C arlo simulations. Management Science, 1981

    1981

  33. [42]

    Sarah Lewis, Tim Hempel, José Jiménez-Luna, Michael Gastegger, Yu Xie, Andrew Y. K. Foong, Victor García Satorras, Osama Abdin, Bastiaan S. Veeling, Iryna Zaporozhets, Yaoyi Chen, Soojung Yang, Adam E. Foster, Arne Schneuing, Jigyasa Nigam, Federico Barbero, Vincent Stimper, A...

    2025

  34. [43]

    Out of many, one: designing and scaffolding proteins at the scale of the structural universe with Genie 2

    Yeqing Lin, Minji Lee, Zhao Zhang, and Mohammed AlQuraishi. Out of many, one: designing and scaffolding proteins at the scale of the structural universe with Genie 2 . arXiv preprint arXiv:2405.15489, 2024 a

    2024 arXiv

  35. [44]

    Nguyen, and Mohammed AlQuraishi

    Yeqing Lin, Haewon C. Nguyen, and Mohammed AlQuraishi. In-silico protein design pipeline . github.com/aqlaboratory/insilico_design_pipeline, 2024 b

    2024

  36. [45]

    Evolutionary-scale prediction of atomic-level protein structure with a language model

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Salvatore Candido, and Alexander Rives. Evolutionary-scale prediction of atomic-level p...

    2023

  37. [46]

    SGDR : Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR : Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016

    2016 arXiv

  38. [47]

    Assessing generative model coverage of protein structures with SHAPES

    Tianyu Lu, Melissa Liu, Yilin Chen, Jinho Kim, and Po-Ssu Huang. Assessing generative model coverage of protein structures with SHAPES . bioRxiv, 2025

    2025

  39. [48]

    Shintaro Minami. PyDSSP . github.com/ShintaroMinami/PyDSSP, 2023

    2023

  40. [49]

    M onte C arlo gradient estimation in machine learning

    Shakir Mohamed, Mihaela Rosca, Michael Figurnov, and Andriy Mnih. M onte C arlo gradient estimation in machine learning. Journal of Machine Learning Research, 2020

    2020

  41. [50]

    Scalable end-to-end autonomous vehicle testing via rare-event simulation

    Matthew O'Kelly, Aman Sinha, Hongseok Namkoong, Russ Tedrake, and John C Duchi. Scalable end-to-end autonomous vehicle testing via rare-event simulation. Advances in Neural Information Processing Systems, 2018

    2018

  42. [51]

    Stochastic Differential Equations: An Introduction with Applications

    Bernt Oksendal. Stochastic Differential Equations: An Introduction with Applications. Springer Science & Business Media, 2013

    2013

  43. [52]

    Empirical likelihood

    Art B Owen. Empirical likelihood. Chapman and Hall/CRC, 2001

    2001

  44. [53]

    gender-guesser

    Israel Saeta P \'e rez, David Arcos, and LeadRatings contributors. gender-guesser. github.com/lead-ratings/gender-guesser, 2016

    2016

  45. [54]

    Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods

    John Platt. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In Advances in Large Margin Classifiers, 1999

    1999

  46. [55]

    Empirical likelihood and general estimating equations

    Jin Qin and Jerry Lawless. Empirical likelihood and general estimating equations. The Annals of Statistics, 1994

    1994

  47. [56]

    Class-balancing diffusion models

    Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, and Ya Zhang. Class-balancing diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  48. [57]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 2023

    2023

  49. [58]

    Black box variational inference

    Rajesh Ranganath, Sean Gerrish, and David Blei. Black box variational inference. In Artificial Intelligence and Statistics. PMLR, 2014

    2014

  50. [59]

    Variational inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, 2015

    2015

  51. [60]

    Convex Analysis

    R Tyrrell Rockafellar. Convex Analysis. Princeton University Press, 1970

    1970

  52. [61]

    SAXS ensemble refinement of ESCRT-III CHMP3 conformational transitions

    Bartosz R \'o \.z ycki, Young C Kim, and Gerhard Hummer. SAXS ensemble refinement of ESCRT-III CHMP3 conformational transitions. Structure, 2011

    2011

  53. [62]

    Designing DNA with tunable regulatory activity using score-entropy discrete diffusion

    Anirban Sarkar, Yijie Kang, Nirali Somia, Pablo Mantilla, Jessica Lu Zhou, Masayuki Nagai, Ziqi Tang, Chris Zhao, and Peter Koo. Designing DNA with tunable regulatory activity using score-entropy discrete diffusion. bioRxiv, 2024

    2024

  54. [63]

    A tutorial on conformal prediction

    Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction. Journal of Machine Learning Research, 2008

    2008

  55. [64]

    Finetuning text-to-image diffusion models for fairness

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In International Conference on Learning Representations, 2024

    2024

  56. [65]

    Waman, Paul Ashford, Harry M

    Ian Sillitoe, Nicola Bordin, Natalie Dawson, Vaishali P. Waman, Paul Ashford, Harry M. Scholes, Camilla S. M. Pang, Laurel Woodridge, Clemens Rauer, Neeladri Sen, Mahnaz Abbasian, Sean Le Cornu , Su Datt Lam, Karel Berka, Ivana Huta r ov \'a Varekov \'a , Radka Svobodova, Jon ...

    2021

  57. [66]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021

    2021

  58. [67]

    Density estimation by dual ascent of the log-likelihood

    Esteban G Tabak and Eric Vanden-Eijnden. Density estimation by dual ascent of the log-likelihood. Communications in Mathematical Sciences, 2010

    2010

  59. [68]

    Fine-tuning of diffusion models via stochastic control: entropy regularization and beyond

    Wenpin Tang. Fine-tuning of diffusion models via stochastic control: entropy regularization and beyond. arXiv preprint arXiv:2403.06279, 2024

    2024

  60. [69]

    Diffusion probabilistic modeling of protein backbones in 3D for the motif-scaffolding problem

    Brian L Trippe, Jason Yim, Doug Tischer, David Baker, Tamara Broderick, Regina Barzilay, and Tommi Jaakkola. Diffusion probabilistic modeling of protein backbones in 3D for the motif-scaffolding problem. In International Conference on Learning Representations, 2023

    2023

  61. [70]

    Fine-tuning of continuous-time diffusion models as entropy-regularized control

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Fine-tuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194, 2024

    2024 arXiv

  62. [71]

    Asymptotic statistics

    Aad W Van der Vaart. Asymptotic statistics. Cambridge University Press, 2000

    2000

  63. [72]

    Graphical models, exponential families, and variational inference

    Martin J Wainwright and Michael I Jordan. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 2008

    2008

  64. [73]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Conference on Computer Vision and Pattern Recognition. IEEE C...

    2024

  65. [74]

    Watson, David Juergens, Nathaniel R

    Joseph L. Watson, David Juergens, Nathaniel R. Bennett, Brian L. Trippe, Jason Yim, Helen E. Eisenach, Woody Ahern, Andrew J. Borst, Robert J. Ragotte, Lukas F. Milles, Basile I. M. Wicky, Nikita Hanikel, Samuel J. Pellock, Alexis Courbet, William Sheffler, Jue Wang, Preetham ...

    2023

  66. [75]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning

    Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 1992

    1992

  67. [76]

    Enforcing statistical constraints in generative adversarial networks for modeling chaotic dynamical systems

    Jin-Long Wu, Karthik Kashinath, Adrian Albert, Dragos Chirila, and Heng Xiao. Enforcing statistical constraints in generative adversarial networks for modeling chaotic dynamical systems. Journal of Computational Physics, 2020

    2020

  68. [77]

    Guide your favorite protein sequence generative model

    Junhao Xiong, Hunter Nisonoff, Maria Lukarska, Ishan Gaur, Luke M Oltrogge, David F Savage, and Jennifer Listgarten. Guide your favorite protein sequence generative model. arXiv preprint arXiv:2505.04823, 2025

    2025

  69. [78]

    SE(3) diffusion model with application to protein backbone generation

    Jason Yim, Brian L Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola. SE(3) diffusion model with application to protein backbone generation. In International Conference on Machine Learning, 2023

    2023

  70. [79]

    Normalizing flows are capable generative models

    Shuangfei Zhai, Ruixiang Zhang, Preetum Nakkiran, David Berthelot, Jiatao Gu, Huangjie Zheng, Tianrong Chen, Miguel \'A ngel Bautista, Navdeep Jaitly, and Joshua M Susskind. Normalizing flows are capable generative models. In International Conference on Machine Learning, 2025

    2025

  71. [80]

    Brookes, Akosua Busia, Ana Carneiro, Clara Fannjiang, Galina Popova, David Shin, Kevin C

    Danqing Zhu, David H. Brookes, Akosua Busia, Ana Carneiro, Clara Fannjiang, Galina Popova, David Shin, Kevin C. Donohue, Li F. Lin, Zachary M. Miller, Evan R. Williams, Edward F. Chang, Tomasz J. Nowakowski, Jennifer Listgarten, and David V. Schaffer. Optimal trade-off control...

    2024

  72. [81]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  73. [82]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  74. [83]

    Generative AI and Biology (GenBio)

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    1982 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.