Diffusion Models are Evolutionary Algorithms

Benedikt Hartl; Hananel Hazan; Michael Levin; Yanbo Zhang

arxiv: 2410.02543 · v3 · submitted 2024-10-03 · 💻 cs.NE · cs.LG

Diffusion Models are Evolutionary Algorithms

Yanbo Zhang , Benedikt Hartl , Hananel Hazan , Michael Levin This is my paper

Pith reviewed 2026-05-23 20:18 UTC · model grok-4.3

classification 💻 cs.NE cs.LG

keywords diffusion modelsevolutionary algorithmsdenoising processoptimizationselectionmutationlatent space

0 comments

The pith

Diffusion models inherently perform evolutionary algorithms by treating evolution as a denoising process.

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

The paper shows that diffusion models are equivalent to evolutionary algorithms. It frames evolution as a denoising process whose reverse is diffusion, which mathematically includes the operations of selection, mutation, and reproductive isolation. From this equivalence the authors derive Diffusion Evolution, a method that uses iterative denoising to refine candidate solutions in parameter space. The method finds multiple optima and outperforms standard evolutionary algorithms on the tasks examined. They further introduce Latent Space Diffusion Evolution, which applies latent diffusion and accelerated sampling to handle high-dimensional problems with fewer steps.

Core claim

By considering evolution as a denoising process and reversed evolution as diffusion, we mathematically demonstrate that diffusion models inherently perform evolutionary algorithms, naturally encompassing selection, mutation, and reproductive isolation. Building on this equivalence, we propose the Diffusion Evolution method: an evolutionary algorithm utilizing iterative denoising to heuristically refine solutions in parameter spaces. Unlike traditional approaches, Diffusion Evolution efficiently identifies multiple optimal solutions and outperforms prominent mainstream evolutionary algorithms. Furthermore, leveraging advanced concepts from diffusion models, namely latent space diffusion and加速

What carries the argument

The equivalence that models evolution as a denoising process (with diffusion as its reverse), which directly supplies the functional roles of selection, mutation, and reproductive isolation inside a diffusion model.

If this is right

Standard diffusion models can be used directly as evolutionary algorithms for optimization without additional training.
Diffusion Evolution locates multiple distinct optima rather than converging to a single solution.
Latent-space and accelerated-sampling techniques from diffusion models reduce the number of steps needed for evolutionary search in high-dimensional spaces.
Non-Gaussian or discrete diffusion formulations can be substituted into the same evolutionary framework.
The equivalence supplies a concrete route for transferring techniques between diffusion-based generative modeling and evolutionary computation.

Where Pith is reading between the lines

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

If the equivalence holds, biological evolutionary dynamics could be simulated by running diffusion models forward or backward in time.
Hybrid algorithms could alternate between diffusion steps and explicit evolutionary operators to gain advantages from both.
The mapping suggests that questions about open-ended evolution might be studied by varying the noise schedule or the form of the diffusion process.
Representations of genomes or phenotypes could be embedded in the latent space of a diffusion model to enable evolutionary search at reduced computational cost.

Load-bearing premise

Modeling evolution as a denoising process and its reverse as diffusion produces a mathematically valid equivalence that captures the functional roles of selection, mutation, and reproductive isolation without material distortion.

What would settle it

A direct comparison on standard benchmark optimization tasks in which the trajectories or final solution sets produced by a trained diffusion model diverge from those produced by an evolutionary algorithm that implements the same selection and mutation rules.

Figures

Figures reproduced from arXiv: 2410.02543 by Benedikt Hartl, Hananel Hazan, Michael Levin, Yanbo Zhang.

**Figure 1.** Figure 1: Evolution processes can be viewed as the inverse process of diffusion, where higher fitness populations [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: (a) Diffusion Evolution on a two-peak fitness landscape: Populations near the two black crosses have higher fitness. For each individual xt (black star), its target xˆ0 (red dots) is estimated by a weighted average of its neighbors (c.f., dots within the blue disks, respectively); larger dot-size indicates higher fitness. The individual then moves a small step forward to the next generation (orange star). … view at source ↗

**Figure 3.** Figure 3: One of the benchmark experiments (columns) on different fitness functions with selected evolutionary algo [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: (a) Evolution process of cart-pole tasks: The horizontal axis shows survival time, and the vertical axis represents generations. Each point indicates an individual’s state (pole angle, cart shift) at their final survival. As the evolution progresses, more systems survive longer and achieve higher rewards. (b) Compared to the original Diffusion Evolution (blue), the Latent Space Diffusion Evolution method (… view at source ↗

**Figure 2.** Figure 2: αt = 1 − 1 T . (14) The second is the schedule used in DDPM, which can be approximated by: αt = exp −β0t − γt2 T , (15) where β0 and γ are hyperparameters. These are calculated by constraining α0 = 1 − ε and αT = ε, with ε = 10−4 as the default. The third schedule is the cosine schedule proposed by Nichol & Dhariwal (2021) and is used in both Figures 3 and 4: αt = 1 2 cos πt T + 1 2 . (16) For σt, … view at source ↗

read the original abstract

In a convergence of machine learning and biology, we reveal that diffusion models are evolutionary algorithms. By considering evolution as a denoising process and reversed evolution as diffusion, we mathematically demonstrate that diffusion models inherently perform evolutionary algorithms, naturally encompassing selection, mutation, and reproductive isolation. Building on this equivalence, we propose the Diffusion Evolution method: an evolutionary algorithm utilizing iterative denoising -- as originally introduced in the context of diffusion models -- to heuristically refine solutions in parameter spaces. Unlike traditional approaches, Diffusion Evolution efficiently identifies multiple optimal solutions and outperforms prominent mainstream evolutionary algorithms. Furthermore, leveraging advanced concepts from diffusion models, namely latent space diffusion and accelerated sampling, we introduce Latent Space Diffusion Evolution, which finds solutions for evolutionary tasks in high-dimensional complex parameter space while significantly reducing computational steps. This parallel between diffusion and evolution not only bridges two different fields but also opens new avenues for mutual enhancement, raising questions about open-ended evolution and potentially utilizing non-Gaussian or discrete diffusion models in the context of Diffusion Evolution.

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

The paper reframes diffusion models as evolutionary algorithms via a denoising-evolution mapping, but the claimed mathematical demonstration looks like a high-level re-description rather than a derivation of selection, mutation, and isolation from the diffusion equations.

read the letter

The central claim is that treating evolution as denoising and its reverse as diffusion shows diffusion models inherently run evolutionary algorithms. The authors build Diffusion Evolution on this, using iterative denoising to refine parameter-space solutions, and add a latent-space version that borrows accelerated sampling to cut steps in high dimensions. They report it finds multiple optima and beats standard evolutionary algorithms. The new pieces are the specific Diffusion Evolution heuristic and the latent-space extension that imports diffusion tricks for optimization efficiency. That cross-field borrowing is the part worth testing if the mapping holds up in practice. The soft spots are exactly where the stress-test note flags them. The abstract asserts a mathematical demonstration but supplies no equations, score-function derivations, or population-level accounting for fitness or reproductive isolation. Without those steps it is hard to tell whether the operators emerge from the noise schedule or are imposed externally in the proposed algorithm. The circularity risk is real: once evolution is defined as denoising, the equivalence follows by construction rather than revealing an independent isomorphism. Performance claims also sit unsupported in the given text. This is for readers already working at the generative-model and evolutionary-optimization boundary who want a new heuristic to try. It is not ready for broad citation until the derivations and benchmarks are visible. I would send it to peer review because the topic is timely and the proposed method could still be useful even if the big equivalence needs tightening.

Referee Report

2 major / 1 minor

Summary. The paper claims that diffusion models are evolutionary algorithms. By framing evolution as a denoising process and reversed evolution as diffusion, it asserts a mathematical demonstration that diffusion models inherently encompass selection, mutation, and reproductive isolation. Building on this, the authors propose Diffusion Evolution, an EA that uses iterative denoising to refine solutions in parameter space, claiming it efficiently identifies multiple optima and outperforms mainstream EAs; they further introduce Latent Space Diffusion Evolution leveraging latent diffusion and accelerated sampling for high-dimensional tasks.

Significance. If the claimed equivalence can be established rigorously without circularity, the work could bridge diffusion-based generative modeling and evolutionary computation, enabling new hybrid algorithms and raising questions about open-ended evolution. The empirical claims of superior performance in locating multiple solutions and reduced computational steps in the latent variant would be of interest if supported by detailed derivations and experiments.

major comments (2)

[Abstract] Abstract: The central claim of a 'mathematical demonstration' that diffusion models 'inherently perform evolutionary algorithms' naturally encompassing selection, mutation, and reproductive isolation is unsupported by any equations, derivations, or explicit mapping from the diffusion score function or noise schedule to these operators.
[Abstract] Abstract: The equivalence is introduced by defining evolution as denoising (and its reverse as diffusion), which creates a risk that subsequent claims are tautological by construction rather than independently derived; no demonstration is supplied showing how population-level fitness or reproductive isolation emerges from the diffusion process itself rather than being imposed externally in the proposed heuristic.

minor comments (1)

[Abstract] The abstract refers to outperforming 'prominent mainstream evolutionary algorithms' without naming specific baselines or providing references.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We address each point below and will make revisions to improve clarity and support for the claims.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim of a 'mathematical demonstration' that diffusion models 'inherently perform evolutionary algorithms' naturally encompassing selection, mutation, and reproductive isolation is unsupported by any equations, derivations, or explicit mapping from the diffusion score function or noise schedule to these operators.

Authors: The abstract summarizes the contribution at a high level. The explicit mapping from the diffusion score function and noise schedule to the evolutionary operators is derived in Sections 2 and 3 of the manuscript, where the reverse process is shown to implement selection via the data likelihood gradient, mutation via the forward noise schedule, and reproductive isolation via independent denoising paths. We will revise the abstract to include a brief parenthetical reference to these derivations for improved support. revision: yes
Referee: [Abstract] Abstract: The equivalence is introduced by defining evolution as denoising (and its reverse as diffusion), which creates a risk that subsequent claims are tautological by construction rather than independently derived; no demonstration is supplied showing how population-level fitness or reproductive isolation emerges from the diffusion process itself rather than being imposed externally in the proposed heuristic.

Authors: The initial framing uses the denoising view as an interpretive bridge motivated by the shared iterative refinement structure, but the demonstration proceeds by showing that the standard diffusion equations independently produce the EA properties: population-level fitness emerges from the score matching the data distribution gradient, and reproductive isolation from separate stochastic trajectories. The Diffusion Evolution algorithm is a downstream heuristic, not the basis of the equivalence. We will revise the abstract to clarify this distinction and add a short derivation paragraph in the introduction. revision: yes

Circularity Check

1 steps flagged

Central equivalence introduced by definitional ansatz on evolution as denoising

specific steps

self definitional [Abstract]
"By considering evolution as a denoising process and reversed evolution as diffusion, we mathematically demonstrate that diffusion models inherently perform evolutionary algorithms, naturally encompassing selection, mutation, and reproductive isolation."

The demonstration is predicated on the definitional premise that evolution equals denoising. Once that premise is adopted, the claim that diffusion models perform evolutionary algorithms (including the three operators) is true by the initial stipulation rather than by deriving population-level fitness, selection, or isolation from the score function or noise schedule.

full rationale

The paper's strongest claim rests on an initial modeling choice that directly supplies the desired conclusion. By stipulating that evolution is a denoising process (and its reverse is diffusion), the subsequent assertion that diffusion models 'inherently perform evolutionary algorithms' and 'naturally encompass' selection, mutation, and reproductive isolation follows by construction rather than from an independent derivation of those operators from the diffusion equations. No external benchmark or non-tautological mapping is shown in the provided text; the functional roles are therefore re-described rather than derived. This matches the self-definitional pattern and justifies a score of 7 (one load-bearing definitional step that renders the central demonstration circular).

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The core modeling choice is the denoising-evolution mapping.

axioms (1)

domain assumption Evolution can be represented as a denoising process whose reverse is diffusion.
This premise is required for the claimed mathematical equivalence and is stated in the abstract.

pith-pipeline@v0.9.0 · 5696 in / 1070 out tokens · 27929 ms · 2026-05-23T20:18:36.907471+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.

By considering evolution as a denoising process and reversed evolution as diffusion, we mathematically demonstrate that diffusion models inherently perform evolutionary algorithms, naturally encompassing selection, mutation, and reproductive isolation.
IndisputableMonolith/Foundation/LogicAsFunctionalEquation.lean SatisfiesLawsOfLogic 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.

xhat0(xt, alpha, t) = 1/Z sum g[f(x)] N(xt; sqrt(alpha t) x, 1-alpha t) x ... sigma t w (mutation term)

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.

Forward citations

Cited by 2 Pith papers

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

Regret Analysis of Guided Diffusion for Black-Box Optimization over Structured Inputs
stat.ML 2026-05 unverdicted novelty 8.0

A certificate-based regret analysis framework for guided-diffusion black-box optimization is introduced, with mass lift as the central quantity explaining convergence from pretrained generators.
A Unification of Discrete, Gaussian, and Simplicial Diffusion
cs.LG 2025-12 unverdicted novelty 6.0

Discrete, Gaussian, and simplicial diffusion models for sequences are unified as parameterizations of the Wright-Fisher population genetics model, allowing multi-domain training and stable simplicial diffusion.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · cited by 2 Pith papers · 6 internal anchors

[1]

The origin of species, 1859-1959

Charles Darwin. The origin of species, 1859-1959. Bios, 30(2):67–72,

work page 1959
[2]

Density estimation using Real NVP

Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803,

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Genetic algorithms in search, optimization, and machine learning

David E Golberg. Genetic algorithms in search, optimization, and machine learning. Addion wesley, 1989(102):36,

work page 1989
[4]

The CMA Evolution Strategy: A Tutorial

Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Classifier-Free Diffusion Guidance

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Extensions of lipshitz mapping into hilbert space

William B Johnson. Extensions of lipshitz mapping into hilbert space. InConference modern analysis and probability, 1984, pp. 189–206,

work page 1984
[7]

Auto-Encoding Variational Bayes

Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

The genomic code: The genome instantiates a generative model of the organism

Kevin J Mitchell and Nick Cheney. The genomic code: The genome instantiates a generative model of the organism. arXiv preprint arXiv:2407.15908,

work page arXiv
[9]

Evolution Strategies as a Scalable Alternative to Reinforcement Learning

Tim Salimans, Jonathan Ho, Xi Chen, Szymon Sidor, and Ilya Sutskever. Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint arXiv:1703.03864,

work page internal anchor Pith review Pith/arXiv arXiv
[10]

Denoising Diffusion Implicit Models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020a. Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020b. Richard S Sutton an...

work page internal anchor Pith review Pith/arXiv arXiv 2010
[11]

Evolutionary algorithms: A critical review and its future prospects

Pradnya A Vikhar. Evolutionary algorithms: A critical review and its future prospects. In 2016 International confer- ence on global trends in signal processing, information computing and communication (ICGTSPICC), pp. 261–265. IEEE,

work page 2016
[12]

Neural network diffusion

Kai Wang, Zhaopan Xu, Yukun Zhou, Zelin Zang, Trevor Darrell, Zhuang Liu, and Yang You. Neural network diffusion. arXiv preprint arXiv:2402.13144,

work page arXiv
[13]

And the σ = 0.1

and µ2 = (−1, −1). And the σ = 0.1. A.2 Alpha and Noise Schedule In our experiments, we tested three different noise schedules for αt. The first is a simple linear schedule, used in Figure 2: αt = 1 − 1 T . (14) The second is the schedule used in DDPM, which can be approximated by: αt = exp −β0t − γt2 T , (15) where β0 and γ are hyperparameters. These are...

work page 2021

[1] [1]

The origin of species, 1859-1959

Charles Darwin. The origin of species, 1859-1959. Bios, 30(2):67–72,

work page 1959

[2] [2]

Density estimation using Real NVP

Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803,

work page internal anchor Pith review Pith/arXiv arXiv

[3] [3]

Genetic algorithms in search, optimization, and machine learning

David E Golberg. Genetic algorithms in search, optimization, and machine learning. Addion wesley, 1989(102):36,

work page 1989

[4] [4]

The CMA Evolution Strategy: A Tutorial

Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772,

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

Classifier-Free Diffusion Guidance

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598,

work page internal anchor Pith review Pith/arXiv arXiv

[6] [6]

Extensions of lipshitz mapping into hilbert space

William B Johnson. Extensions of lipshitz mapping into hilbert space. InConference modern analysis and probability, 1984, pp. 189–206,

work page 1984

[7] [7]

Auto-Encoding Variational Bayes

Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

The genomic code: The genome instantiates a generative model of the organism

Kevin J Mitchell and Nick Cheney. The genomic code: The genome instantiates a generative model of the organism. arXiv preprint arXiv:2407.15908,

work page arXiv

[9] [9]

Evolution Strategies as a Scalable Alternative to Reinforcement Learning

Tim Salimans, Jonathan Ho, Xi Chen, Szymon Sidor, and Ilya Sutskever. Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint arXiv:1703.03864,

work page internal anchor Pith review Pith/arXiv arXiv

[10] [10]

Denoising Diffusion Implicit Models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020a. Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020b. Richard S Sutton an...

work page internal anchor Pith review Pith/arXiv arXiv 2010

[11] [11]

Evolutionary algorithms: A critical review and its future prospects

Pradnya A Vikhar. Evolutionary algorithms: A critical review and its future prospects. In 2016 International confer- ence on global trends in signal processing, information computing and communication (ICGTSPICC), pp. 261–265. IEEE,

work page 2016

[12] [12]

Neural network diffusion

Kai Wang, Zhaopan Xu, Yukun Zhou, Zelin Zang, Trevor Darrell, Zhuang Liu, and Yang You. Neural network diffusion. arXiv preprint arXiv:2402.13144,

work page arXiv

[13] [13]

And the σ = 0.1

and µ2 = (−1, −1). And the σ = 0.1. A.2 Alpha and Noise Schedule In our experiments, we tested three different noise schedules for αt. The first is a simple linear schedule, used in Figure 2: αt = 1 − 1 T . (14) The second is the schedule used in DDPM, which can be approximated by: αt = exp −β0t − γt2 T , (15) where β0 and γ are hyperparameters. These are...

work page 2021