arxiv: 2605.11485 · v1 · submitted 2026-05-12 · 💻 cs.RO

Recognition: 2 theorem links

· Lean Theorem

Coordinated Diffusion: Generating Multi-Agent Behavior Without Multi-Agent Demonstrations

Lasse Peters , Laura Ferranti , Javier Alonso-Mora , Andrea Bajcsy

Authors on Pith no claims yet

Pith reviewed 2026-05-13 02:04 UTC · model grok-4.3

classification 💻 cs.RO

keywords multi-agent imitation learningdiffusion policiescoordinated behaviorcost-guided samplingsingle-agent demonstrationsrobot coordination

0 comments

The pith

Coordinated Diffusion couples single-agent diffusion policies with a cost function to produce joint multi-agent behavior without any coordinated training data.

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

The paper demonstrates a way to generate coordinated actions among multiple agents, such as robot arms, by training separate diffusion models on individual agent demonstrations and then steering their combination during sampling. A user-defined multi-agent cost function supplies the coordination signal through an added guidance term in the diffusion score, which can be computed without gradients or extra training. This approach addresses the exponential growth in required data as the number of agents increases, because joint demonstrations become unnecessary once the single-agent behaviors are covered. If the composition works, coordinated policies become feasible with far less expensive data collection focused on solo tasks.

Core claim

CoDi decomposes the diffusion sampling score into the sum of independent single-agent base policies plus a cost-driven guidance term; the guidance steers the product of the base policies toward desired joint outcomes and can be estimated in a gradient-free way for black-box costs. The method succeeds when single-agent data covers the support of the target multi-agent distribution and the cost selects the right combinations from that product.

What carries the argument

The cost-driven guidance term added to the decomposed diffusion score, which coordinates independent single-agent policies into joint behavior during sampling.

If this is right

Multi-agent policies can be learned from single-agent demonstrations plus a cost function rather than joint data.
The guidance term works with non-differentiable black-box costs without requiring additional training.
Data efficiency improves because exponential growth in joint state-action space is avoided.
The composition faithfully approximates target behavior only when base-policy support is sufficient and the cost is well-designed.

Where Pith is reading between the lines

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

The same decomposition might allow cost-guided coordination in other generative models besides diffusion.
For more than two agents, the approach would require costs that remain tractable as the product space grows.
Hardware validation beyond two-arm tasks could test whether real-world dynamics preserve the theoretical coverage condition.

Load-bearing premise

Single-agent demonstrations must contain every behavior needed for the target joint distribution, and the cost function must be able to promote the desired combinations from the product of those independent policies.

What would settle it

Train single-agent policies on data that deliberately omits key coordinated configurations required by the target task, then run sampling with the cost function; the generated joint trajectories will fail to satisfy the coordination objective.

Figures

Figures reproduced from arXiv: 2605.11485 by Andrea Bajcsy, Javier Alonso-Mora, Lasse Peters, Laura Ferranti.

**Figure 1.** Figure 1: Coordinated Diffusion (CoDi) is a method for generating multi-agent behavior from single-agent demonstrations. Abstract—Imitation learning powered by generative models has proven effective for modeling complex single-agent behaviors. However, teaching multi-agent systems, like multiple arms or vehicles, to coordinate through imitation learning is hindered by a fundamental data bottleneck: as the joint stat… view at source ↗

**Figure 2.** Figure 2: Distribution of initial cube positions used for training. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative example of multi-agent behavior generated by CoDi. Our approach guides the left robot to pick up the cube [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison with baselines that utilize multi-agent [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Performance of CoDi and classifier guidance under [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 7.** Figure 7: Comparison with vanilla classifier guidance as alterna [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 6.** Figure 6: Comparison with independent single-agent guidance. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 8.** Figure 8: Comparison against variants that use a simplified cost [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

read the original abstract

Imitation learning powered by generative models has proven effective for modeling complex single-agent behaviors. However, teaching multi-agent systems, like multiple arms or vehicles, to coordinate through imitation learning is hindered by a fundamental data bottleneck: as the joint state-action space grows exponentially with the number of agents, collecting a sufficient amount of coordinated multi-agent demonstrations becomes extremely costly. In this work, we ask: how can we leverage single-agent demonstration data to learn multi-agent policies? We present Coordinated Diffusion (CoDi), a framework that couples independently trained single-agent diffusion policies through a user-defined multi-agent cost function, without requiring any coordinated demonstrations. We derive a new diffusion-based sampling scheme wherein the diffusion score function decomposes into independent, single-agent pre-trained base policies plus a cost-driven guidance term that coordinates these base policies into cohesive multi-agent behavior. We show that this guidance term can be estimated in a gradient-free manner, making CoDi applicable to black-box, non-differentiable cost functions without additional training. Theoretically and empirically, we analyze the conditions under which this composition can faithfully approximate a target multi-agent behavior. We find a complementary role for demonstration data versus the cost function: single-agent demonstrations must cover the support of the desired multi-agent behavior, while the cost function must promote desired behavior from this product of single-agent policies. Our results in simulation and hardware experiments of a two-arm manipulation task show that CoDi discovers robust coordinated behavior from single-agent data, is more data-efficient than multi-agent baselines, and highlights the importance of joint guidance, base policy support, and cost design.

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.

Desk Editor's Note private letter to a colleague

CoDi shows how to steer pre-trained single-agent diffusion policies into coordinated multi-agent behavior with a user cost function and no joint data, but only when the single-agent support already covers the needed states.

read the letter

CoDi takes independently trained single-agent diffusion policies and adds a guidance term derived from a user-defined multi-agent cost to produce coordinated behavior without any joint demonstrations. That is the main new piece: a diffusion sampling scheme where the joint score is the sum of the individual base scores plus the cost-driven correction, estimated gradient-free so it works with black-box costs. They also lay out the supporting condition that single-agent data must cover the support of the target joint distribution while the cost only reweights within it. The two-arm manipulation results in simulation and on hardware demonstrate that the approach can generate coordinated motions and is more data-efficient than the multi-agent baselines they compare against. Credit for running the hardware experiments and for being explicit about the data-versus-cost division of labor. The math on the score decomposition looks clean from the description, and the gradient-free estimation is a practical plus. The soft spot is the support condition. The paper states it and analyzes it, but the reported experiments do not include controlled cases that shrink the base-policy support while holding the cost fixed to show where the composition breaks. Without those boundary tests it is harder to know how far the method extends beyond the two-arm setup where the single-agent data already overlaps heavily with the coordinated states. No other major gaps stand out in the abstract and stress-test details. This is for people doing imitation learning on robots who already have single-agent data and can define a cost for coordination, such as dual-arm or vehicle teams. A reader interested in diffusion models or multi-agent robotics would get a usable recipe plus the right caveats. I would send it to peer review; the core idea is grounded enough and the honesty about assumptions makes it worth referee time even if the empirical coverage of the limits could be tighter.

Referee Report

2 major / 1 minor

Summary. The paper introduces Coordinated Diffusion (CoDi), a framework that generates multi-agent coordinated behaviors by coupling independently trained single-agent diffusion policies via a user-defined multi-agent cost function, without any multi-agent demonstrations. It derives a diffusion sampling scheme in which the joint score decomposes into the sum of the pre-trained single-agent scores plus a cost-driven guidance term, shows that the guidance can be estimated gradient-free, and analyzes (theoretically and empirically) the conditions under which the composition approximates a target joint distribution. The key condition identified is that single-agent data must cover the support of the desired multi-agent behavior while the cost shapes the product distribution; results are shown on two-arm manipulation in simulation and hardware, claiming greater data efficiency than multi-agent baselines.

Significance. If the decomposition and support conditions hold as analyzed, the work would meaningfully advance imitation learning for multi-agent robotics by removing the exponential cost of collecting joint demonstrations. The gradient-free guidance for black-box costs and the explicit complementary roles of data support versus cost are useful contributions. The approach is falsifiable via the stated support-coverage requirement and could be extended to other generative models, though its practical scope is bounded by the need for single-agent data to already include the relevant coordinated states.

major comments (2)

[Experiments] The central claim that the composition faithfully approximates the target multi-agent distribution rests on the support-coverage condition (single-agent demonstrations must cover the support of desired joint behavior). The abstract states this condition is analyzed theoretically and empirically, yet the two-arm experiments do not include controlled ablations that deliberately shrink base-policy support (e.g., by removing demonstrations of specific relative poses while holding the cost fixed) to test the boundary case where guidance cannot recover missing probability mass. This omission weakens the empirical validation of the load-bearing theoretical condition.
[Method] The derived sampling scheme equates the joint score to the sum of independent single-agent scores plus the guidance term derived from the cost. This identity is exact if and only if the target joint is proportional to the product of the marginals times the cost factor. The manuscript should provide the explicit derivation (including any regularity assumptions on the diffusion process) in the methods section, as any unstated approximation in the score decomposition directly affects whether the guidance term can coordinate the policies as claimed.

minor comments (1)

[Abstract] Notation for the guidance term and the product distribution could be introduced earlier and used consistently to improve readability of the decomposition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review. We appreciate the identification of areas where the manuscript can be strengthened and address each major comment below.

read point-by-point responses

Referee: [Experiments] The central claim that the composition faithfully approximates the target multi-agent distribution rests on the support-coverage condition (single-agent demonstrations must cover the support of desired joint behavior). The abstract states this condition is analyzed theoretically and empirically, yet the two-arm experiments do not include controlled ablations that deliberately shrink base-policy support (e.g., by removing demonstrations of specific relative poses while holding the cost fixed) to test the boundary case where guidance cannot recover missing probability mass. This omission weakens the empirical validation of the load-bearing theoretical condition.

Authors: We agree that the current experiments primarily demonstrate success under sufficient support coverage and do not include explicit negative ablations testing the boundary where support is deliberately reduced. While our theoretical analysis and existing results (showing coordination only when relevant states are present in single-agent data) support the condition, adding controlled ablations would provide stronger empirical validation. In the revised manuscript, we will include new experiments that systematically remove demonstrations of specific relative poses from the single-agent datasets (while keeping the cost fixed) and demonstrate that the guidance term fails to recover the missing coordinated behaviors, directly testing the support-coverage requirement. revision: yes
Referee: [Method] The derived sampling scheme equates the joint score to the sum of independent single-agent scores plus the guidance term derived from the cost. This identity is exact if and only if the target joint is proportional to the product of the marginals times the cost factor. The manuscript should provide the explicit derivation (including any regularity assumptions on the diffusion process) in the methods section, as any unstated approximation in the score decomposition directly affects whether the guidance term can coordinate the policies as claimed.

Authors: We thank the referee for this clarification request. The explicit derivation of the score decomposition (showing that the joint score equals the sum of single-agent scores plus the cost-derived guidance term exactly when the target joint distribution is proportional to the product of the single-agent marginals modulated by the cost) is provided in the appendix, along with the regularity assumptions (standard Gaussian diffusion forward process with known score function and the target being defined via the product measure times the cost factor). We will move this full derivation, including the stated assumptions, into the main Methods section of the revised manuscript to make the exactness of the identity and its implications for coordination explicit. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation is a standard score decomposition under explicitly stated support assumption

full rationale

The central derivation applies the known log-gradient identity to a product distribution (joint score = sum of marginal scores + cost gradient), which is a mathematical fact independent of the paper's data or claims. The paper explicitly flags the support-coverage requirement as a necessary condition rather than deriving it from the method itself, and the cost function is user-defined and external rather than fitted to multi-agent targets. No self-citations are invoked to justify uniqueness or load-bearing premises, and no fitted parameters are relabeled as predictions. The result is therefore self-contained against external diffusion-model theory.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the diffusion score admits an additive decomposition into independent single-agent terms plus a guidance term derived from an arbitrary cost; no new physical entities are postulated.

axioms (1)

domain assumption The diffusion score function for the joint multi-agent process can be expressed as the sum of independent single-agent scores plus a cost-driven guidance term.
This decomposition is the load-bearing step in the new sampling scheme described in the abstract.

pith-pipeline@v0.9.0 · 5593 in / 1311 out tokens · 47870 ms · 2026-05-13T02:04:09.815522+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

the diffusion score function decomposes into independent, single-agent pre-trained base policies plus a cost-driven guidance term... π(a|s) := 1/Z(s) exp(−J(s,a)/λ) p1:N_θ(a|s)
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

single-agent demonstrations must cover the support of the desired multi-agent behavior, while the cost function must promote desired behavior from this product of single-agent policies

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

42 extracted references · 42 canonical work pages · 7 internal anchors

[1]

Diffusion policy: Vi- suomotor policy learning via action diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burch- fiel, R. Tedrake, and S. Song, “Diffusion policy: Vi- suomotor policy learning via action diffusion,”Proc. of Robotics: Science and Systems (RSS), 2023

work page 2023
[2]

Diffusion policy: Visuomotor policy learning via action diffusion,

——, “Diffusion policy: Visuomotor policy learning via action diffusion,”Intl. Journal of Robotics Research (IJRR), vol. 44, no. 10-11, pp. 1684–1704, 2025

work page 2025
[3]

Planning with Diffusion for Flexible Behavior Synthesis

M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” arXiv preprint, 2022, arXiv:2205.09991

work page internal anchor Pith review arXiv 2022
[4]

Goal- conditioned imitation learning using score-based diffu- sion policies,

M. Reuss, M. Li, X. Jia, and R. Lioutikov, “Goal- conditioned imitation learning using score-based diffu- sion policies,”Proc. of Robotics: Science and Systems (RSS), 2023

work page 2023
[5]

Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collab- oration 0,

A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jainet al., “Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collab- oration 0,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 6892–6903

work page 2024
[6]

DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset

A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karamcheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Elliset al., “Droid: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[7]

Opti- mal and approximate q-value functions for decentralized pomdps,

F. A. Oliehoek, M. T. Spaan, and N. Vlassis, “Opti- mal and approximate q-value functions for decentralized pomdps,”Journal of Artificial Intelligence Research, vol. 32, pp. 289–353, 2008

work page 2008
[8]

Deep unsupervised learning using nonequi- librium thermodynamics,

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequi- librium thermodynamics,” inProc. of the Int. Conf. on Machine Learning (ICML). Proceedings of Machine Learning Research, 2015, pp. 2256–2265

work page 2015
[9]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 6840–6851, 2020

work page 2020
[10]

Generative modeling by es- timating gradients of the data distribution,

Y . Song and S. Ermon, “Generative modeling by es- timating gradients of the data distribution,”Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019

work page 2019
[11]

Score-based generative mod- eling through stochastic differential equations,

Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative mod- eling through stochastic differential equations,”Proc. of the Int. Conf. on Learning Representations (ICLR), 2021

work page 2021
[12]

Elucidating the design space of diffusion-based generative models,

T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” Advances in Neural Information Processing Systems (NeurIPS), vol. 35, pp. 26 565–26 577, 2022

work page 2022
[13]

Stochastic interpolants: A unifying framework for flows and diffusions,

M. Albergo, N. M. Boffi, and E. Vanden-Eijnden, “Stochastic interpolants: A unifying framework for flows and diffusions,”Journal of Machine Learning Research, vol. 26, no. 209, pp. 1–80, 2025

work page 2025
[14]

Diffusion meets flow matching: Two sides of the same coin,

R. Gao, E. Hoogeboom, J. Heek, V . D. Bortoli, K. P. Murphy, and T. Salimans, “Diffusion meets flow matching: Two sides of the same coin,” 2024. [Online]. Available: https://diffusionflow.github.io/

work page 2024
[15]

Flow matching for generative modeling,

Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inProc. of the Int. Conf. on Learning Representations (ICLR), 2023

work page 2023
[16]

Diffusion models beat gans on image synthesis,

P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in Neural Information 12 Processing Systems (NeurIPS), vol. 34, pp. 8780–8794, 2021

work page 2021
[17]

Diffwave: A versatile diffusion model for audio synthesis,

Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Diffwave: A versatile diffusion model for audio synthesis,” inInternational Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview.net/forum?id=a-xFK8Ymz5J

work page 2021
[18]

Video diffusion models,

J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. Fleet, “Video diffusion models,” inAdvances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associates, Inc., 2022, pp. 8633–8646. [Online]. Available: https: //proceedings.neurips.cc/paper files/paper/2022/file/...

work page 2022
[19]

Diffusion probabilistic models for 3d point cloud generation,

S. Luo and W. Hu, “Diffusion probabilistic models for 3d point cloud generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 2837–2845

work page 2021
[20]

MIMIC- D: Multi-modal imitation for MultI-agent coordination with decentralized diffusion policies,

D. Dong, M. Bhatt, S. Choi, and N. Mehr, “MIMIC- D: Multi-modal imitation for MultI-agent coordination with decentralized diffusion policies,”arXiv preprint arXiv:2509.14159, 2025

work page internal anchor Pith review arXiv 2025
[21]

Latent theory of mind: A decentralized diffusion architecture for cooperative manipulation,

C. He, G. S. Camps, X. Liu, M. Schwager, and G. Sar- toretti, “Latent theory of mind: A decentralized diffusion architecture for cooperative manipulation,”arXiv preprint arXiv:2505.09144, 2025

work page arXiv 2025
[22]

Compositional visual generation with composable diffu- sion models,

N. Liu, S. Li, Y . Du, A. Torralba, and J. B. Tenenbaum, “Compositional visual generation with composable diffu- sion models,” inProceedings of the European Conference on Computer Vision (ECCV). Springer, 2022, pp. 423– 439

work page 2022
[23]

Reduce, reuse, recycle: Compositional gen- eration with energy-based diffusion models and MCMC,

Y . Du, C. Durkan, R. Strudel, J. B. Tenenbaum, S. Diele- man, R. Fergus, J. Sohl-Dickstein, A. Doucet, and W. S. Grathwohl, “Reduce, reuse, recycle: Compositional gen- eration with energy-based diffusion models and MCMC,” inProceedings of the International Conference on Ma- chine Learning (ICML). PMLR, 2023, pp. 8489–8510

work page 2023
[24]

PoCo: Policy composition from and for heterogeneous robot learning,

L. Wang, J. Zhao, Y . Du, E. H. Adelson, and R. Tedrake, “PoCo: Policy composition from and for heterogeneous robot learning,” inRobotics: Science and Systems (RSS), 2024

work page 2024
[25]

Multi-robot motion planning with diffusion models,

Y . Shaoul, I. Mishani, S. Vats, J. Li, and M. Likhachev, “Multi-robot motion planning with diffusion models,” in International Conference on Learning Representations (ICLR), 2025

work page 2025
[26]

Efficient online reinforcement learning for diffusion policy,

H. Ma, T. Chen, K. Wang, N. Li, and B. Dai, “Efficient online reinforcement learning for diffusion policy,” in International Conference on Machine Learning. PMLR, 2025, pp. 41 837–41 853

work page 2025
[27]

EXPO: Stable Reinforcement Learning with Expressive Policies

P. Dong, Q. Li, D. Sadigh, and C. Finn, “Expo: Stable reinforcement learning with expressive policies,”arXiv preprint arXiv:2507.07986, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[28]

Diffusion policy policy optimization,

A. Z. Ren, J. Lidard, L. L. Ankile, A. Simeonov, P. Agrawal, A. Majumdar, B. Burchfiel, H. Dai, and M. Simchowitz, “Diffusion policy policy optimization,” inInternational Conference on Learning Representations (ICLR), 2025

work page 2025
[29]

Diffusing to coordinate: Efficient online multi-agent diffusion policies,

Z. Li, H. Zhong, X. Wang, Q. Xia, L. Zhang, and L. Huang, “Diffusing to coordinate: Efficient online multi-agent diffusion policies,”arXiv preprint arXiv:2602.18291, 2026

work page arXiv 2026
[30]

MADiff: Offline multi-agent learning with diffusion models,

Z. Zhu, M. Liu, L. Mao, B. Kang, M. Xu, Y . Yu, S. Ermon, and W. Zhang, “MADiff: Offline multi-agent learning with diffusion models,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

work page 2024
[31]

Classifier-Free Diffusion Guidance

J. Ho and T. Salimans, “Classifier-free diffusion guid- ance,”arXiv preprint arXiv:2207.12598, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[32]

Hierarchical Text-Conditional Image Generation with CLIP Latents

A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image genera- tion with clip latents,”arXiv preprint arXiv:2204.06125, vol. 1, no. 2, p. 3, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[33]

The Principles of Diffusion Models,

C.-H. Lai, Y . Song, D. Kim, Y . Mitsufuji, and S. Ermon, “The principles of diffusion models,”arXiv preprint, 2025, arXiv:2510.21890

work page arXiv 2025
[34]

Reverse-time diffusion equation mod- els,

B. D. Anderson, “Reverse-time diffusion equation mod- els,”Stochastic Processes and their Applications, vol. 12, no. 3, pp. 313–326, 1982

work page 1982
[35]

A connection between score matching and denoising autoencoders,

P. Vincent, “A connection between score matching and denoising autoencoders,”Neural computation, vol. 23, no. 7, pp. 1661–1674, 2011

work page 2011
[36]

Asymptotic eval- uation of certain markov process expectations for large time, i,

M. D. Donsker and S. S. Varadhan, “Asymptotic eval- uation of certain markov process expectations for large time, i,”Communications on pure and applied mathe- matics, vol. 28, no. 1, pp. 1–47, 1975

work page 1975
[37]

R. B. Nelsen,An introduction to copulas. Springer, 2006

work page 2006
[38]

Fonctions de r ´epartition `a n dimensions et leurs marges,

M. Sklar, “Fonctions de r ´epartition `a n dimensions et leurs marges,” inAnnales de l’ISUP, vol. 8, no. 3, 1959, pp. 229–231

work page 1959
[39]

Vector copulas,

Y . Fan and M. Henry, “Vector copulas,”Journal of Econometrics, vol. 234, no. 1, pp. 128–150, 2023

work page 2023
[40]

An empirical bayes approach to statis- tics,

H. E. Robbins, “An empirical bayes approach to statis- tics,” inBreakthroughs in Statistics: Foundations and basic theory. Springer Verlag, 1992, pp. 388–394

work page 1992
[41]

Tweedie’s formula and selection bias,

B. Efron, “Tweedie’s formula and selection bias,”Journal of the American Statistical Association, vol. 106, no. 496, pp. 1602–1614, 2011

work page 2011
[42]

Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning

V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. All- shire, A. Handaet al., “Isaac gym: High performance gpu-based physics simulation for robot learning,”arXiv preprint, 2021, arXiv:2108.10470. 13 APPENDIXA DERIVATIONS A. Optimally Compensating Multi-Agent Cost Function For completeness, we begin by repeatin...

work page internal anchor Pith review Pith/arXiv arXiv 2021