Denoising Diffusion-based Evolutionary Algorithm

• DDEA is defined as an evolutionary framework where offspring are produced through controlled denoising dynamics, replacing traditional mutation and crossover.
• It integrates fitness-weighted Bayesian estimation and online diffusion training to guide candidate solutions in both continuous and combinatorial optimization tasks.
• Empirical evaluations show DDEA achieves competitive fitness and maintains diversity across multimodal, high-dimensional, and dynamic search landscapes.

Denoising Diffusion-based Evolutionary Algorithm (DDEA) denotes a class of evolutionary methods in which a denoising diffusion process functions as an explicit search operator over candidate solutions rather than merely as a generative model. In current formulations, diffusion appears in at least three closely related roles: as the entire evolutionary update derived directly from a fitness-weighted denoising process, as an online-trained variation operator over continuous genotypes, and as a learned initializer and recombination operator inside a conventional evolutionary algorithm for combinatorial optimization (Zhang et al., 2024, Hartl et al., 2024, Soler et al., 8 Oct 2025). In generative settings, the same perspective extends to explicit crossover in diffusion trajectory space, where the initial noise and per-timestep reverse-process noise sequence are treated as the genotype and offspring are produced by controlled interpolation during denoising (Kumada et al., 16 Apr 2026).

1. Taxonomic scope and defining characteristics

A DDEA is not a single algorithmic template but a family of methods unified by one defining move: offspring are produced through denoising dynamics. In the continuous-parameter setting, HADES replaces Gaussian mutation, covariance adaptation, and crossover with a trainable denoising diffusion model trained generation by generation on historical genotypes and fitnesses, with high-fitness samples up-weighted (Hartl et al., 2024). In the combinatorial setting, the DDEA framework for the Maximum Independent Set Problem uses pre-trained denoising diffusion models for diffusion-based population initialization and a conditional diffusion recombination operator trained by imitation learning against an optimized recombination demonstrator (Soler et al., 8 Oct 2025). In the most direct theoretical construction, Diffusion Evolution dispenses with neural denoiser training entirely and implements denoising through a population-based Bayesian estimator built from fitness-weighted kernels (Zhang et al., 2024).

Formulation	Core diffusion role	Domain
Diffusion Evolution	Fitness-weighted denoising update over a population	Continuous optimization, RL
HADES	Online-trained diffusion variation operator	Continuous parameter search
CHARLES-D	Conditional diffusion variation with classifier-free guidance	Continuous search, RL
DDEA for MIS	Diffusion-based initialization and parent-aware recombination	Combinatorial optimization
Diffusion crossover	Explicit crossover over reverse-process noise sequences	Interactive image evolution

Across these variants, several features recur. Candidate solutions are represented directly in the diffusion model’s working space, whether as continuous vectors $x \in \mathbb{R}^D$, binary incidence vectors relaxed to node-score heatmaps, or reverse-process noise trajectories. Selection pressure is injected either through a fitness-to-density mapping, through a fitness-weighted denoising loss, or through elitist parent selection and demonstrator-guided recombination. Mutation is realized by stochasticity in reverse diffusion, by forward noising followed by denoising, or by fresh noise at non-interpolated timesteps. A plausible implication is that DDEA is best understood as a representationally flexible meta-framework in which evolutionary operators are re-expressed as controlled diffusion transformations.

2. Diffusion–evolution equivalence

The most explicit theoretical basis for DDEA is the claim that diffusion models are evolutionary algorithms when a fitness function is mapped to a probability density $p(x)=g[f(x)]$, with $g(\cdot)$ monotone (Zhang et al., 2024). Under standard diffusion notation, the forward process is

$$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$

and reverse denoising iteratively transforms an initial Gaussian population into samples from the high-fitness density. In this reading, forward diffusion corresponds to reverse evolution, because it destroys structure; reverse diffusion corresponds to forward evolution, because it denoises random parameters toward well-adapted regions.

The crucial step is a Bayesian estimator for the latent “ancestor” $x_0$ of a noisy individual $x_t$. Using the fitness-defined prior and the Gaussian corruption model, the posterior mean estimator becomes

$$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$

with

$$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$

This estimator defines selection and recombination simultaneously. Higher-fitness individuals receive larger weights through $g[f(x)]$, while the Gaussian kernel enforces locality, so only sufficiently nearby individuals contribute strongly. The resulting denoising step is written in DDIM-style form as

$$x_{t-1} = \sqrt{\alpha_{t-1}}\,\hat{x}_0 + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\,\hat{\epsilon} + \sigma_t\,\epsilon,$$

where

$p(x)=g[f(x)]$0

In this formulation, the population at diffusion time $p(x)=g[f(x)]$1 is the evolutionary population at generation $p(x)=g[f(x)]$2; the $p(x)=g[f(x)]$3 term is mutation; the weighted posterior mean is selection plus soft many-parent recombination; and the shrinking kernel neighborhood induces reproductive isolation around distinct optima. The paper’s latent extension, Latent Space Diffusion Evolution, performs the neighborhood computation in a random-projection latent space $p(x)=g[f(x)]$4 while still evaluating fitness in parameter space, thereby mitigating high-dimensional distance concentration (Zhang et al., 2024).

3. Online diffusion strategies in continuous parameter spaces

HADES instantiates DDEA as an evolutionary strategy whose variation operator is a denoising diffusion model trained and adapted online from evolutionary data (Hartl et al., 2024). Candidate solutions are continuous genotypes $p(x)=g[f(x)]$5, and the diffusion model operates directly in parameter space rather than in a separate latent embedding. The forward process is standard:

$p(x)=g[f(x)]$6

while the denoiser $p(x)=g[f(x)]$7 is trained with a diffusion loss modified by fitness weighting:

$p(x)=g[f(x)]$8

The weighting function $p(x)=g[f(x)]$9 biases the learned generative density toward high-fitness regions. The paper states that this rescaling is equivalent to drawing training data from a density $g(\cdot)$0, so the learned density approximates $g(\cdot)$1. Reverse sampling is DDIM-like, with stochasticity controlled by $g(\cdot)$2.

The evolutionary loop differs from a classical selection–reproduction scheme. Instead of explicitly selecting parents and then applying fixed mutation and crossover, HADES updates a buffer $g(\cdot)$3 containing historical genotypes, fitnesses, and optional conditions; fine-tunes the diffusion model on this buffer every generation; and then samples the next population by reverse denoising from Gaussian seeds or from proto-genotypes constructed by elite crossover. The paper’s pseudocode includes optional mutation via a forward diffusion step of scale $g(\cdot)$4 and optional partial readaptation via denoising for $g(\cdot)$5 steps before full offspring generation. In this sense, mutation is the stochastic part of diffusion sampling, recombination is optional crossover before denoising, and selection is embodied in buffer composition and the weighting $g(\cdot)$6.

CHARLES-D extends HADES by introducing classifier-free guidance at the parameter level. The denoiser becomes $g(\cdot)$7, where $g(\cdot)$8 may encode genotypic traits, target fitness, novelty, or phenotypic traits in RL. The target fitness can be sampled in “Fisher-style” as $g(\cdot)$9 or in “Greedy” form as $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$0. Novelty can be derived from k-NN distances, and RL phenotypes can be conditioned through tuples such as $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$1. Guided sampling combines conditional and unconditional predictions in the usual classifier-free form, with guidance strength $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$2 controlling the bias toward the target condition. The authors interpret the buffer plus diffusion model as a form of deep memory that retains historical structure across generations (Hartl et al., 2024).

4. Learned diffusion operators for combinatorial optimization

The combinatorial DDEA framework of Salva Soler and colleagues applies the paradigm to the Maximum Independent Set Problem on Erdős–Rényi graphs $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$3 with $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$4, using solution encodings $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$5, where $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$6 indicates membership in the independent set (Soler et al., 8 Oct 2025). The framework targets ER-50-100, ER-300-400, ER-700-800, and an out-of-distribution ER-1300-1400 regime. Because diffusion outputs are continuous node-score vectors rather than directly feasible independent sets, a decoder sorts vertices by descending $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$7, greedily inserts non-ineligible vertices into the set, and marks neighbors ineligible. This decoder is used for diffusion-based initialization, diffusion-based recombination, and mutation.

This DDEA preserves a standard EA backbone: initialization, binary tournament selection, recombination, mutation, and elitist replacement. Its novelty lies in replacing low-information operators with learned diffusion operators. First, a pre-trained DIFUSCO-style graph diffusion model generates $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$8 initialization heatmaps $$x_t = \sqrt{\alpha_t}\,x_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon \sim \mathcal{N}(0,I),$$9, sampled in parallel and decoded into feasible individuals. The model uses an anisotropic GNN with edge gating, hidden dimension $x_0$0, $x_0$1 layers, Gaussian diffusion for ER graphs, $x_0$2 training steps, and $x_0$3 denoising steps at inference. It is trained on MIS labels from KaMIS with a 60s time limit.

Second, the framework introduces a conditional diffusion recombination operator trained by imitation learning against an optimized recombination demonstrator. The demonstrator is an ILP that, given two parents $x_0$4 and $x_0$5, maximizes child cardinality subject to MIS feasibility and a Hamming-distance budget $x_0$6, where $x_0$7 and $x_0$8. Solving this ILP with Gurobi under a 15s time limit yields expert children $x_0$9, and approximately $x_t$0 triplets $x_t$1 are collected across the three ER ranges. The conditional diffusion model is then trained to denoise noisy versions of $x_t$2 conditioned on the graph and both parent solutions. At runtime, DDEA uses binary tournament selection, diffusion recombination, mutation with $x_t$3 and deselection probability $x_t$4, elitist replacement, and a non-duplication rule (Soler et al., 8 Oct 2025).

The significance of this construction is methodological. It shows that DDEA need not mean online re-training from a rolling archive; it can also mean embedding pre-trained diffusion inference inside the EA loop as a high-quality initializer and a parent-aware learned recombination operator. The paper explicitly describes the framework as intended to be problem-independent up to two problem-specific components: a decoder from probabilities to feasible solutions and a mechanism for generating training labels.

5. Recombination, memory, and controllability

Recombination is the most conceptually variable component across DDEA formulations. In Diffusion Evolution, recombination is implicit: each individual’s denoised target $x_t$5 is a fitness-weighted average over many nearby individuals, so hybridization occurs through the posterior mean itself (Zhang et al., 2024). In HADES, recombination may occur before denoising by forming crossover seeds $x_t$6, after which the reverse process “repairs” the seed into a plausible high-fitness genotype; the paper explicitly likens this to diffusion inpainting (Hartl et al., 2024). In the MIS framework, recombination is learned directly as a conditional distribution $x_t$7 over offspring conditioned on two parent solutions (Soler et al., 8 Oct 2025).

The most explicit operator-level definition appears in Diffusion crossover, which treats the genotype of an image as $x_t$8, where $x_t$9 is the initial Gaussian noise and $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$0 is the sequence of reverse-process noise injections (Kumada et al., 16 Apr 2026). Given parents $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$1 and $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$2, crossover is defined by spherical linear interpolation (Slerp) of both the initial noise and selected per-timestep noise vectors. For one timestep $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$3,

$$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$4

where $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$5 is the angle between the parent noise vectors. Interpolation is applied only over a controlled range of early reverse steps, $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$6; for later steps, fresh Gaussian noise is sampled. The parameter $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$7 controls parent bias, while $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$8 controls how much of the reverse trajectory is inherited rather than mutated.

This design yields a principled exploration–exploitation mechanism. Large $$\hat{x}_0(x_t,\boldsymbol\alpha,t) = \frac{1}{Z} \sum_{x \in \mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big)\,x,$$9 keeps offspring trajectories close to the between-parents manifold and thus emphasizes exploitation; small $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$0 delegates more timesteps to fresh noise and thus emphasizes exploration. The paper reports monotonic LPIPS behavior as $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$1 varies and a monotonic decrease in offspring diversity as $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$2 increases, with strongly negative Spearman correlations on both MNIST and ModelNet. This suggests that DDEA recombination can be defined at several representational levels—parameter vectors, learned conditional distributions, or reverse-process noise sequences—without leaving the diffusion framework. That plurality is one of the field’s distinguishing methodological features.

6. Empirical behavior, strengths, and unresolved constraints

Empirical findings across current DDEA variants are heterogeneous but consistently oriented around multi-modality, controllability, and search efficiency. Diffusion Evolution attains fitness $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$3 on Rosenbrock, Beale, and Himmelblau while maintaining elite-population entropy values $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$4, $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$5, and $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$6, respectively; on Ackley it reaches fitness $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$7 with entropy $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$8, whereas CMA-ES attains entropy $$Z= \sum_{x\in\mathcal{P}_t} g[f(x)]\, \mathcal{N}\big(x_t;\sqrt{\alpha_t}x,\,1-\alpha_t\big).$$9 but only fitness $g[f(x)]$0; on Rastrigin it achieves fitness $g[f(x)]$1 with entropy $g[f(x)]$2, while CMA-ES attains entropy $g[f(x)]$3 but fitness $g[f(x)]$4 (Zhang et al., 2024). In cart-pole, the latent variant is reported to solve the task within only about 10 generations with 512 individuals, including a $g[f(x)]$5-parameter controller.

In the online continuous-parameter setting, HADES tracks the alternating optimum on the dynamic double-peak landscape, whereas CMA-ES often gets “stuck” around one peak and cannot follow the changing optimum. On static double-peak tasks with quadrant conditioning, CHARLES-D converges to the peak located in the conditioned quadrant even though both peaks are symmetric in fitness. On Rastrigin and twisted-Rastrigin, HADES and novelty-conditional CHARLES-D reach global optima faster than CMA-ES and SimpleGA, particularly in the twisted case. Novelty guidance yields high entropy and discovery of multiple optima, on average about 3 out of 4 peaks and sometimes all 4 within a single population. In cart-pole RL, HADES solves the task in about 3–4 generations with $g[f(x)]$6, outperforming CMA-ES and SimpleGA by about an order of magnitude in generations, and phenotype-level conditioning can bias balancing behavior toward specified resting cart positions without changing the reward function (Hartl et al., 2024).

In combinatorial optimization, the MIS DDEA consistently outperforms DIFUSCO under matched time budgets. On ER-300-400, DDEA-long attains cost $g[f(x)]$7 versus $g[f(x)]$8 for DIFUSCO-timed, with shorter runtime and fewer diffusion calls; on ER-700-800, it attains $g[f(x)]$9 versus $$x_{t-1} = \sqrt{\alpha_{t-1}}\,\hat{x}_0 + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\,\hat{\epsilon} + \sigma_t\,\epsilon,$$0, again with slightly less time. Under approximately the same time budget, DDEA-long’s independent sets are 3.9% larger than Gurobi’s on ER-300-400 and 7.5% larger on ER-700-800. In out-of-distribution evaluation on ER-1300-1400, DDEA reaches cost $$x_{t-1} = \sqrt{\alpha_{t-1}}\,\hat{x}_0 + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\,\hat{\epsilon} + \sigma_t\,\epsilon,$$1 versus $$x_{t-1} = \sqrt{\alpha_{t-1}}\,\hat{x}_0 + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\,\hat{\epsilon} + \sigma_t\,\epsilon,$$2 for DIFUSCO-timed, corresponding to 11.6% higher quality under comparable runtime. Ablations attribute gains to both diffusion-based initialization and diffusion-based recombination, with the latter having the larger effect (Soler et al., 8 Oct 2025).

In image-generation experiments, Diffusion crossover produces offspring trajectories that lie between parent trajectories in PCA space, visually smooth transitions as $$x_{t-1} = \sqrt{\alpha_{t-1}}\,\hat{x}_0 + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\,\hat{\epsilon} + \sigma_t\,\epsilon,$$3 varies, and decreasing LPIPS diversity as $$x_{t-1} = \sqrt{\alpha_{t-1}}\,\hat{x}_0 + \sqrt{1-\alpha_{t-1}-\sigma_t^2}\,\hat{\epsilon} + \sigma_t\,\epsilon,$$4 increases. Interactive experiments reported convergence toward targets such as a “thick/bold 5” and a “cushioned L-shaped sofa,” with global geometry stabilizing while local texture continues to vary (Kumada et al., 16 Apr 2026). These results do not constitute a general-purpose optimization benchmark, but they do establish that explicit diffusion-space crossover can support coherent human-in-the-loop evolution.

The principal strengths attributed to DDEA across the literature are expressive non-Gaussian variation, multimodal search, archive-like or trajectory-level memory, and conditional controllability. The principal constraints are equally clear. HADES notes compute overhead from training a diffusion model every generation, sensitivity to buffer size, selection pressure, diffusion schedules, mutation/readaptation times, guidance strength, and architecture, and unresolved scalability to very high-dimensional problems (Hartl et al., 2024). The MIS framework depends on expensive offline demonstrator generation, problem-specific decoders, and GPU-intensive diffusion inference, though the authors note that the current implementation has inefficient in-GPU parallelization (Soler et al., 8 Oct 2025). Diffusion crossover requires storing full noise sequences, currently assumes the system itself generated the parents, and has only qualitative IEC evidence rather than broader user studies (Kumada et al., 16 Apr 2026). A plausible overall assessment is that DDEA has already diversified into continuous, combinatorial, and generative forms, but its long-term status will depend on whether the added representational power and controllability continue to offset the substantial algorithmic and computational overheads observed in current implementations.