Generative Models in Decision Making: A Survey

Yinchuan Li, Xinyu Shao, Jianping Zhang, Haozhi Wang, Leo Maxime Brunswic,Kaiwen Zhou, Jiqian Dong, Kaiyang Guo, Xiu Li, Zhitang Chen, Jun Wang, Jianye Hao

Abstract—In recent years, the exceptional performance of generative models in generative tasks has sparked significant interest in their integration into decision-making processes. Due to their ability to handle complex data distributions and their strong model capacity, generative models can be effectively incorporated into decision-making systems by generating trajectories that guide agents toward high-reward state-action regions or intermediate sub-goals. This paper presents a comprehensive review of the application of generative models in decision-making tasks. We classify seven fundamental types of generative models: energy-based models, generative adversarial networks, variational autoencoders, normalizing flows, diffusion models, generative flow networks, and autoregressive models. Regarding their applications, we categorize their functions into three main roles: controllers, modelers and optimizers, and discuss how each role contributes to decision-making. Furthermore, we examine the deployment of these models across five critical real-world decision-making scenarios. Finally, we summarize the strengths and limitations of current approaches and propose three key directions for advancing next-generation generative directive models: high-performance algorithms, large-scale generalized decision-making models, and self-evolving and adaptive models.

Index Terms—Generative Models, Decision Making, Generative Decision Making

1 INTRODUCTION

ENERATIVE models have become a hot topic in both G academia and industry, primarily due to their ability to generate large quantities of synthetic data with high quality and diversity. From early systems like DALL-E [1] (for image generation) and GPT-3 [2] (for text generation) to more recent advancements such as DALL-E3 [3], ChatGPT and GPT-4 [4], generative models have rapidly advanced in both quality and scale of their outputs.

Content generation aims to create coherent material that mimics training examples, while decision-making focuses on producing action sequences for optimal outcomes. Unlike content generation, decision-making involves complex, dynamic environments and long-term decisions. Thus, despite generative models’ success in content generation, applying them to decision-making poses challenges. These challenges include: 1) how to learn policies through interaction with the environment, rather than simply mimicking expert behavior. 2) how to generate new policies based on learned behaviors, transitioning from policy learning to policy generation. 3) how to establish a robust fundamental decision generating model that can adapt to various environments with minimal tuning efforts. 4) how to build multi-step reasoning and long-term evolution capabilities of strategies. These challenges emphasize the need for generative models to go beyond mere data generation.

In practice, decision-making is often referred to as sequential decision making, where a decision maker makes a series of observations over time, and each decision influencs subsequent choices. The goal is to identify a policy that optimizes expected rewards or minimizes costs across sequential actions. Classical algorithms such as Dynamic Programming (DP) and Reinforcement Learning (RL) are widely used to solve problems modeled as Markov Decision Processes (MDPs). These methods optimize decisionmaking by updating policies based on observed rewards and state transitions rather than generating new ones. Despite their many successful applications, these traditional approaches often rely on trial-and-error or predefined states and transitions, limiting exploration and potentially missing better solutions. Moreover, they require substantial computation and optimization, which can be impractical for highdimensional or large-scale problems. Traditional methods also need significant reconfiguration or retraining to new environment, reducing flexibility.

On the other hand, generative models are designed to model data distributions, rather than simply fitting labels. Once trained, they can generate new samples that resemble the original data, enabling the exploration of diverse scenarios and outcomes. This ability enables the discovery of novel strategies that may not be immediately evident with traditional methods. In complex or poorly labeled data scenarios, generative models offer a richer understanding of possible decision paths, sometimes leading to strategies that better align with high rewards or desired goals. However, traditional methods like optimization or reinforcement learning remain effective in simpler, welldefined environments where the decision space is clearer, and goals are more straightforward. The choice between these approaches depends on the task complexity and the environment characteristics.

Recognizing these advantages, recent years have seen substantial research efforts aimed at developing new generative models and applying them to decision making. Fig. 1 illustrates the research trends in generative models and their applications in decision making, further emphasizing the significance of these methods in addressing such challenges. However, there is a lack of comprehensive reviews that summarize past work and pave the way for new research avenues. This gap motivates us to provide this review paper.

Fig. 1: Research trends in generative models and their applications in decision-making (2000-2024). The bars represent the rough average annual number of papers, sourced from Google Scholar. The search included titles associated with seven classic types of generative models and their applications in five real-world scenarios we detailed in Fig. 20.

The survey highlights three major contributions: 1) A comprehensive taxonomy is proposed for classifying current generative decision-making methods. We identify seven types of generative models used in decision-making and categorize their functions into three key roles: controller, modeler, and optimizer. 2) We review the diverse practical uses of generative models in decision-making, focusing on robot control, structural generation, games, autonomous driving and optimization tasks. 3) Finally, we summarize the advantages and limitations of existing work and discuss future perspectives for developing high-performance generative models in decision-making tasks.

The organization of the rest of this survey is as follows (refer to Fig. 2 for a general outline): Section 2 serves as a preliminary by introducing sequential decision-making formulation and provides the basics of all the examined methods. Specifically, we offer a detailed introduction to seven types of generative models and compare their performance with traditional approaches. Section 3 presents the proposed taxonomy for categorizing generative decisionmaking methods. In Section 4, we review and analyze existing literature according to the introduced taxonomy. Section 5 showcases practical applications of generative models in decision-making. Finally, Section 6 discusses the future directions of generative models in decision-making, and we conclude the paper in Section 7 with an overall summary.

2 PRELIMINARIES

2.1 Sequential Decision Making

Sequential decision-making involves making step-by-step decisions, where each choice depends on previous outcomes. At each step, the agent typically observes the current state, selects an optimal action based on its policy, and the environment updates its state while providing new rewards. The ultimate goal is to maximize the accumulated rewards.

The Sequential decision process is categorized according to the number of agents and the observation to the environment. In this paper, we only focus on the single agent. If the agent is able to fully observe the environment, it is considered a Markov Decision Process (MDP) [5]. An MDP is a framework for modeling discrete-time decisionmaking process, where outcomes depend on both random factors and the agent’s decisions [5]. Formally, an MDP is defined as a tuple M = (S, A, R, P, ρ0, γ, H) to represent MDP, where S is the state space, and s ∈ S contains all the information perceived by the agent from the environment. A represents the action space, encompassing all possible actions a that the agent can execute when interacting with the environment. r ∈ R ∶ S × A × S → R denotes the reward function, which corresponds to the transition pair (s, a, s′). P ∶ S × A → ∆(S) refers to the transition function. When the action a is applied to the state s, P (s′∣s, a) generates a distribution over states. ρ0 ∈ ∆(S) represents the initial state distribution. The discount factor γ ∈ [0, 1] quantifies the long-term value of the current action, while H denotes the horizon.

If observations are limited, we can define the Partially Observed Markov Decision Process (POMDP) [6] as a tuple M = (S, A, O, R, P, ρ0, E, γ, H). Here, O denotes the observation space, and each observation ot ∈ O is obtained by the emission function E(ot∣st).

The goal of sequential decision making is to learn a policy π(a∣s) or π(a∣o) by optimizing the expected reward:

$$\tag{1}$$

where pπ(τ ) is the distribution over trajectories τ induced by policy π:

$$\tag{2}$$

2.2 Related Methods

We introduce several traditional methods for solving sequential decision making problems in this section.

2.2.1 Imitation Learning

Imitation learning aims to optimize a policy πθ(s) such that the action distribution is close to that of an expert policy π∗ for any state s, the objective function is given by

$$\tag{3}$$

where p(s∣πθ) denotes the state distribution induced by πθ, and L is a metric function.

2.2.2 Search based Methods

Policy search focuses on exploring the policy space directly, rather than computing the value function explicitly. For a given policy π, the expected discounted return is given by

$$\tag{4}$$

With large state space, U(π) could be approximated by sampling trajectories consists of (S, A, R) pairs and U(π) can be reformulated as [7]:

$$\tag{5}$$

Fig. 2: An overview of the survey. Specific sections are distinguished by different colors. Best viewed in color.

where R(τ ) denotes the discounted return related to τ . Monte Carlo policy evaluation entails estimating the expected utility of the policy π by performing numerous rollouts starting from s0 ∼ ρ0(s). Local search starts with a feasible solution and incrementally moves to the neighbor with better utility over the search space until convergence occurs, e.g. Hooke-Jeeves method [8]. Simulated annealing [9] allows for occasional moves to feasible solutions with worse utility to approximate the global optimum. Genetic algorithms evaluate different simulations concurrently based on the objective, recombine them and guide the population toward a global optimum.

2.2.3 Planning & Optimization

Planning is an optimization approach that relies on a predefined model to guide the search process. Let U(s) denote the action space for each state s, which represents the set of all actions that could be applied from s. For distinct s, s′ ∈ S, U (s) and U (s′) are not necessarily disjoint. As part of the planning problem, a set SG ⊂ S of goal states and the initial state sI are defined.

Let πK denote a K-step plan, which is a sequence of (a1, a2, ⋯, aK) of K actions. Let F denote the final stage where F = K + 1. The cost functional is defined as

$$\tag{6}$$

where l(sk, ak) is the cost term yielding a real value for every sk ∈ S and ak ∈ U(sk). The final term is defined lF (sF ) = 0 if sf ∈ SG and lF (sF ) = ∞ otherwise.

2.2.4 Reinforcement Learning

The field of RL contains various methods that cater to different decision-making tasks, but all standard RL algorithms generally follow the same learning principles. In particular, the agent interacts the current MDP M using some sort of behavior policy, which can be the present policy π(a∣s) or mix it with a random policy. Then the agent can receive the subsequent state st+1 along with the reward function rt. After repeating this process for many steps, the agent can use the collected samples {s, a, r, s′} to update the policy.

Value-based Methods. One classic approach is to directly estimate the value function of the state of state-action so that a near-optimal policy can be obtained. Define the state-action value as

$$\tag{7}$$

and the state value V π(st) as

$$\tag{8}$$

Then the recursive representation of Qπ(st, at) can be derived as [10]:

$$\tag{9}$$

which can also be expressed as the Bellman operator Bπ, i.e., Qπ = BπQπ. Bπ has the unique fixed point obtained by repeating iterations Qπk+1 = BπQπk [11]. Based on this property, we can derive the modern value iteration (VI) algorithms. Q-learning is a common VI method that expresses the policy as π(at∣st) = δ(at = arg maxat Q(st, at)) with Dirac function δ(⋅). The optimal Q-function can be approximated by substituting this policy into (9). To derive a learning algorithm, we can define a parametric Q-function estimator Qϕ(s, a), which can be optimized by minimizing the difference between LHS and RHS in Bellman equation, such as fitted Q-learning [12]. With the help of neural networks, existing deep Q-learning methods can achieve more accurate value-function estimation by minimizing the Bellman error objective [13].

Policy Gradients. Another classic approach is to directly estimate the gradient of (1). We define policy πθ(a∣s) parameterized by θ, which can be a neural network and output the logits of action a. Then, We can then write the gradient of (1) with respect to θ as:

where Aˆ(st, at) is the return estimator, written by

$$\tag{10}$$

which can be calculated with Monte Carlo samples [14]. b(st) denotes the baseline, which can be approximated as the mean return of the collected trajectories or by the value function V (st) [15].

By introducing KL regularization and clipped surrogate objective, True Region Policy Optimization (TRPO) [16] and Proximal Policy Optimization (PPO) [17] constrain policy size updated at each iteration to stable the training process. These methods and their variants have achieved great success in areas like robot control [18], games [19] and generative model training [20].

Actor-Critic Methods. Taking advantage of both policy gradient and value-based approaches, Actor-critic algorithms use a parameterized policy and value function. And the value function can provide a better estimation of Aˆ(st, at) for policy gradients, with calculating the average return over the sampled trajectories.

Unlike Q-learning, the actor-critic method seeks to optimize the Q-function associated with the current policy πθ rather than learning the optimal Q-function directly. Many variants are proposed based on this principle, such as onpolicy and off-policy algorithms [21]. Deep Deterministic Policy Gradient (DDPG) introduces a deterministic policy under the Actor-Critic architecture, which can solve continuous control problems [21]. Soft actor-critic introduce the entropy regularization into the policy optimization process, so that the agent can make better exploration [22].

Model-based RL. For model-based RL, the first step is to learn the environment model P (⋅∣s, a) based on historical trajectories. According to the probabilistic transition model, a common approach is to reduce the KL divergence between the learned model Pˆ(⋅∣s, a) and the true dynamics P (⋅∣s, a) as [23]:

$$\tag{11}$$

where θ denotes the learned model’s parameters. The model learning process can be transformed as a supervised learning task, which can be solved effectively by some supervised learning technique. To reduce the error of model learning, several scholars propose the Lipschitz continuity constraints [24], distribution matching [25], and robust model learning methods [26] for further precise estimation.

When the model is ready, the agent develops plans to improve its strategy to interact with the modeled world. Traditional methods for integrating planning into MBRL mainly include model predictive control, Monte Carlo tree search and Dyna-style methods [27]. Dyna-style methods utilize the learned model to generate more experience and then perform RL on the model-augmented dataset, which has been the mainstream method [28]. Similarly, these methods employ value function estimation and policy gradient to improve the policy.

2.3 Generative Models

Generative Model [29], [30], [31], [32], [33], [34], [35], [36] is an important subdivision of artificial intelligence that employs techniques and models specifically designed for sampling unseen data from the underlying distribution of the existing dataset.

In this section, we explore generative models through three crucial dimensions: sample quality, sample diversity, and computational efficiency [37], [38]. These dimensions are essential for understanding how generative models perform in decision-making, as they directly impact the accuracy, robustness, and practical applicability of the generated outputs. Sample quality measures how well the generated samples align with the real data distribution, reflecting the realism and reliability of the outputs. High sample quality ensures that decisions based on the generated data are both accurate and trustworthy in real-world contexts. Sample diversity, on the other hand, evaluates how well the model can generate a broad range of distinct samples, capturing the full spectrum of potential outcomes. This is vital in decision-making tasks, where diverse options are necessary to adapt to varying conditions and avoid narrow, overfitted solutions. Finally, computational efficiency refers to the computational resources required for training and inference, making it a key consideration for the deployment of generative models in real-world applications. In contexts such as autonomous driving or robotic control, where decisions must be made in real time and under resource constraints, computational efficiency is critical to ensure that the model remains practical and scalable. Together, these dimensions provide a comprehensive framework for evaluating generative models, guiding their application in complex, resource-sensitive decision-making environments.

We will examine each of the seven generative models through these three crucial dimensions, identifying their strengths and weaknesses. Generative models face challenges in balancing sample quality, sample diversity, and computational efficiency. For example, Diffusion Models and Normalizing Flows offer strong sample diversity and stability but require high computational resources, limiting their suitability for real-time decision-making applications [39], [40], [41]. In contrast, models like VAEs and GANs provide faster training and better efficiency but may struggle with maintaining sample diversity, potentially leading to overly similar or overfitted outputs [34], [42], [43], [44]. Based on the references [43], [45], [46], [47], [48], [49], [50], [51], [52], [53], we compare the performance of these seven generative models across sample quality, diversity, and efficiency, as shown in Fig. 3.

Fig. 3: Comparison of seven generative models in decisionmaking: training stability, generation diversity, and computational efficiency. Larger bubbles represent higher computational efficiency, with different models indicated by distinct colors. Best viewed in color.

Energy Based Models (EBMs). An EBM [54] assigns an energy function Eθ(x) to each data x [55], which is the unnormalized log probability of the data to measure the compatibility of x with the model. The density of x is

$$\tag{12}$$

where Zθ denotes the normalizing constant, which also known as partition function:

$$\tag{13}$$

The model is trained to reduce the energy of data point from the training set while increasing the energy of other, possibly unrealistic, data points [27]. The earliest EBM is the Boltzmann machine [33], which is a stochastic version of the Hopfield network [56]. Since the energy function has no restrictions, it can be parameterized freely and can model any high-dimensional complex data [57], [58] like images [59] and natural languages [60]. However, the intractable normalizing constant poses training and sampling challenges. Typical solutions include 1) MCMC sampling [61]; 2) score matching [62] and 3) noise contrastive estimation [63].

Generative Adversarial Networks (GANs). GANs [34] are well-known due to their distinctive architecturewhich involves a generator G that creates synthetic samples and a discriminator D that evaluates them by distinguishing between real and generated data. The target loss of a typical GAN is given by

$$\tag{14}$$

While GAN can generate realistic contents in image synthesis [34], image style transfer [35], and behavior [64], it suffers from training instability and mode collapse. Several variants of GANs have be proposed to solve the problem, like Wasserstein GANs [65], [66], CycleGANs [35], and Progressive GANs [67], which use different loss functions, architectures, or training techniques.

Variational Autoencoders (VAEs). VAEs [43], known for their stable learning process and efficient computation, attract significant attention in recent years. VAEs [68] are probabilistic models that represent the input data in a compressed form, referred to as the latent space. New samples are generated by sampling from this latent space and then decoding them back into the original data domain. The goal of a VAE, presented in (15), aims to maximize the evidence lower bound (ELBO):

$$\tag{15}$$

In (15), the first term represents the likelihood of the observed data x given latent variable z, and the expectation is taken over the distribution of z. The KL term measures the divergence between the variational distribution q(z), which approximates the posterior distribution of z, and the prior distribution p(z).

VAEs are deployed to generate new images or videos [69], and augment images, such as image inpainting [70] and image colorization [71]. VAEs are used to compress data [72], [73] or provide a good way to detect anomalies or outliers in data [74]. However, VAEs may suffer from low generation quality and the independence assumption between the latent variables [72]. To address these limitations, several extensions of VAEs have been proposed, e.g., conditional VAE [75] and Gaussian mixture VAE [76].

Researchers propose to combine VAEs with other techniques, including adversarial training [77], [78] and normalizing flows [79] for improving disentanglement. Besides, [80] and [81] study the continual learning of VAE representations and [82] trains a response function of the hyperparameter β in β −V AE to trade off the reconstruction error and the KL divergence.

Normalizing Flows (NFs). Flow-based models [83] learn to map a simple random variable, such as a Gaussian distribution, to a more complex distribution capable of representing data. Normalizing Flow [84], [85] applys a specific type of invertible transformation to the simple distribution to preserve the tractability of the density function, as shown below:

$$\tag{16}$$

A primary strength of normalizing flows is their precision in capturing complex probability. Normalizing flows can also be used to generate new samples and to perform density estimation, among other tasks. Despite their success, normalizing flows still face some challenges, such as the restricted network structure and the rendered limited expressiveness. Nevertheless, Normalizing flows remain a dynamic field of study, with continuous advancements being made.

From limited pre-neural attempts [86], [87], [88], specific flow transformations [89] are proposed to reduce the computational cost, such as NICE [90], Real-NVP [91], and Glow [41]. Researchers also use autoregressive flows [92], [93], [94] to accelerate the sampling process. Many Normalizing Flow variants also leverage the techniques of Neural ODE [95], [96], [97] and optimal transport [98], [99].

Diffusion Models (DMs). Diffusion models, introduced by Sohl-Dickstein et al. in 2015 [36], learn to generate samples through a diffusion process. This process involves gradually introducing noise to the data, followed by the reversal of this noise addition during sampling to recover the original data (as illustrated in Fig. 4). The key strength of diffusion models lies in their ability to generate highquality, realistic, and diverse outputs, surpassing earlier generative models such as GANs in various image synthesis tasks [100]. This capability has sparked significant interest in their application to decision-making scenarios, particularly in high-dimensional environments.

Fig. 4: A diagrammatic depiction of diffusion models and one denoising step is illustrated [101].

The key components of a diffusion model are the forward and the reverse process. In the forward phase, Gaussian noise is progressively introduced to the data. At time step t, the data point xt−1 is transformed into a noisy version xt according to Equation 17:

$$\tag{17}$$

Here, βt denotes a schedule of noise variance, which governs how quickly the noise is introduced at step t. The forward process effectively destroys the original data structure, making it increasingly difficult to recover the original input as the process progresses.

The sampling process works by reversing the diffusion (or noise addition) steps. The model iteratively refines a noisy sample to recover the original data by learning a reverse diffusion process:

$$\tag{18}$$

The reverse process can be computationally expensive, as it involves many time steps, each requiring careful sampling to accurately reconstruct the data. Despite this challenge, diffusion models can generate high-quality samples, particularly in fields such as image generation, where they have outperformed GANs in terms of visual fidelity and diversity.

In the context of decision-making, diffusion models offer several advantages. They can model complex, highdimensional decision spaces and generate solutions that explore a wide range of potential outcomes, making them suitable for applications in reinforcement learning (RL), robotics, and automated planning.

For instance, in robotic control, diffusion models can generate motion sequences that not only mimic the desired behavior but also explore alternative strategies that could lead to better performance in uncertain or dynamic environments. Similarly, in autonomous driving, they can generate realistic trajectories and responses to dynamic driving scenarios, taking into account the uncertainty of the environment.

Moreover, diffusion models can facilitate the exploration of decision spaces in complex environments where traditional models might struggle to account for the full range of possibilities. This capability is particularly useful in multiobjective decision making, where a model needs to balance trade-offs between conflicting goals.

Diffusion models demonstrate remarkable potential for producing high-quality, realistic, and novel data, and they continue to be an important tool in the field of generative modeling. Nevertheless, diffusion models have a slow sampling process, which will limit its scalability to highdimensional cases. Consistency models [102] are a potential solution that support fast one-step generation.

GFlowNets/CFlowNets. GFlowNets is a new type of generative model, which [51], [103] are initially defined on finite directed acyclic graphs G = (S, E) with given initial and final states s0, sf ∈ S and given reward function R(s → sf ); an edge flow is simply a map F ∶ E → R+. Fig.5 shows the illustration of the structure of GFlowNets. From such an edge flow, one can define a Markov chain (pt) on G F (s→s′ ) is trained so that for all state s ∈ S ∖ {s0, sf }

The first equality is the flow-matching constraint, the second is the reward constraint. Assuming the reward R is nonnegative and not identically zero, those two constraints ensures that the Markov chains reaches sf at some finite time τ and that the last non-terminal position pτ−1 of the Markov chain follows the distribution obtained by normalizing R R(s→sf )
extended beyond graphs and acyclicity [104], [105], [106], with variation on losses and regularizations [105], [107], [108].

Fig. 5: Illustration of the structure of a Generative Flow Network, as a pointed DAG over states s, with particles flowing along edges to represent the flow function [51].

Autoregressive Models (AMs). Autoregressive models are commonly used in natural language processing and generative modeling tasks. They are a key component of sequence-to-sequence models, such as the Transformer model [109], and are responsible for generating output sequences based on an input sequence or a given context.

Autoregressive models operate by predicting each output element sequentially conditioned on the preceding ones. This means that the generation process is sequential, with the model generating the elements in a left-to-right fashion. At each step, the model takes into account the previously generated elements and the context information to make predictions.

A popular method for autoregressive generation is using an autoregressive language model, such as GPT [110], which has proven successful in NLP tasks like text generation, machine translation, and dialogue systems.

In autoregressive models, the input to each step is typically a combination of the previously generated elements and a context vector that captures the overall context or representation of the input sequence. The context vector can be obtained from an encoder network that processes the input sequence and produces a fixed-size representation, which is then used as input to the decoder.

During training, autoregressive models are typically optimized using Maximum Likelihood Estimation (MLE) [111]. The model aims to optimize the likelihood of generating target output sequence given the input sequence or context. This involves calculating the probability of each target element based on the preceding elements and updating parameters to improve its predictions [112].

Overall, autoregressive models are a powerful tool for sequence generation tasks, particularly in natural language processing [113]. They have contributed to significant advancements in generative modeling and continue to be an active area of research in the field.

2.4 Difference with previous approaches

Reinforcement Learning (RL) and generative models are two distinct branches of machine learning that serve different purposes and tackle different challenges. The key difference lies in their primary objectives and the nature of the learning process [114], as show in Fig. 6.

The primary objective of RL is for an agent to learn an optimal policy that maximizes cumulative rewards through trial-and-error with the environment. The agent learns by taking actions, observing the resulting states, then adjusting current strategy according to the returns. On the other hand, generative models aims to understand and represent the latent data distribution in a given dataset [115]. These models are used to generate new samples resembling the training data, capturing inherent patterns and structures in the data.

The second one is the learning process. In RL, the agent learns through trial and error by exploring the environment, taking actions, and receiving rewards. As it gains experience, the agent refines its policy to make more effective decisions, ultimately maximizing its rewards. Meanwhile, generative models learn from a given dataset and analyze the data, attempting to estimate the probability distribution that generates the data, such as VAEs and GANs.

Last but not least, they differ in how they interact with environment [116]. The RL agent interacts with the environment iteratively, learning from feedback and adjusting its behavior without generating any data. In contrast, the generative model, once trained, is capable of producing entirely new data points resembling the original ones [117], thus actively generating rather than merely responding.

In summary, reinforcement learning focuses on learning optimal actions to achieve specific goals [118], whereas generative models aim to model data distributions and generate new, similar samples. Both are crucial in machine learning but address different challenges.

3 TAXONOMY

In this section, we state our taxonomy to group the generative approaches for solving sequential decision making problems. We categorize the methodologies into five key dimensions, outlined as follows. This taxonomy is further illustrated in Table 1.

Fig. 6: A comparative framework: traditional and generative decision-making (MDP-based vs. data-driven).

Family. The first dimension is the family of the generative models. We represent the family of the approach by the acronym of the category of the generative models. To be more specific, Energy Based Models (EBMs), Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Normalizing Flow (NFs), Diffusion Models (DMs), GFlowNets (GFNs), and Autoregressive Models (AMs).

Function. The second dimension functions on the role of generative models in sequential decision making. Because they can model a diverse set of distributions, generative models can serve various roles in addressing sequential decision-making problems. We basically classify the functions into three categories.

Controller. Generative models can serve as policy models to guide the decision-making process by modeling high-reward trajectories or observationaction pairs. They generate candidate actions based on current observations, refining these candidates to achieve optimal decisions. In this context, sampling from learned distributions is a key step that informs the decision-making process.

Modeler. Generative models are capable of capturing the underlying patterns of the dataset and generating new data resembling the original’s distribution. By producing samples that capture the data’s structure, they support tasks like data augmentation, privacy preservation, and simulation, providing rich inputs for subsequent optimization or decision-making.

Optimizer. Generative models can explore highdimensional spaces and optimize solutions by iteratively refining sampled candidates. They generate and evaluate samples to navigate towards better solutions, helping to optimize decision-making processes or high-reward outcomes.

Model	Family	Function	Structure	Expertise	Application
Du & Mordatch [119]	EBM	Controller	EBM	Imitation Learning	Robot Control
EBIL [120]	EBM	Controller	EBM	Imitation Learning	Robot Control
BM[33]	EBM	Modeler	EBM	Generation	Optimization / Structural Generation
DEBMs[121]	EBM	Modeler	EBM	Online RL	Robot Control /Optimization
SGMs [100]	EBM	Modeler	EBM	Generation	Structural Generation
SO-EBM [122]	EBM	Optimizer	EBM	Others	Optimization
pcEBM[123]	EBM	Optimizer	EBM	Generation	Structural Generation / Optimization
	EBM	Optimizer	EBM	Generation
CF-EBM[124]	GAN	Controller	GAN		Structural Generation
GAIL [64]	GAN	Controller	WGAN	Imitation Learning	Robot Control
InfoGAIL [125]	GAN	Controller	GAN	Imitation Learning Imitation Learning	Robot Control
MGAIL [126]] FAGIL [127]	GAN	Controller			Robot Control
	GAN	Controller	WGAN	Imitation Learning	Robot Control
WGAIL [128]			GAN	Imitation Learning	Robot Control
IC-GAIL [129]	GAN	Controller	GAN	Imitation Learning	Robot Control
AIRL [130]	GAN	Controller	GAN	Imitation Learning	Robot Control
AugAIRL [131]	GAN	Controller	GAN	Imitation Learning	Robot Control
WAIL [132]	GAN	Controller	WGAN	Imitation Learning	Robot Control
MAIRL [133]	GAN	Controller	GAN	Imitation Learning	Robot Control
EGAN[134]	GAN	Modeler	GAN	Online RL	Structural Generation
S2P[135]	GAN	Modeler	GAN	Offline RL	Robot Control
He et al. [136]	GAN	Optimizer	GAN	Generation	Optimization /Games
DCGANs [137]	GAN	Optimizer	DCGAN	Generation	Optimization
C-GANs [138]	GAN	Optimizer	C-GAN	Generation	Optimization
GTI[139]	VAE	Controller	CVAE	Imitation Learning	Robot Control
HULC[140]	VAE	Controller	seq2seq CVAE	Imitation Learning	Robot Control
Play-LMP[141]	VAE	Controller	seq2seq CVAE	Imitation Learning	Robot Control
TACO-RL [142]	VAE	Controller	seq2seq CVAE	Imitation Learning	Robot Control
OPAL [143]	VAE	Controller	βVAE	Imitation Learning	Structural Generation
MaskDP[144]	VAE	Controller	MAE	Offline RL	Robot Control
Han &Kim[145]	VAE	Modeler	VAE	Offline RL	Structural Generation
CageBO [146]	VAE	Optimizer	CVAE	Others	Optimization
CVAE-Opt[147]	VAE	Optimizer	CVAE	Others	Optimization
NF Policy [148]	NF	Controller	Coupling Flow	Offline RL	Robot Control
CNF[149]	NF	Controller	Autoregressive Flow	Offline RL	Autonomous Driving
Guided Flows[150]	NF	Controller	Continuous Flow	Offline RL /Generation	Optimization /Games
NICE[90]	NF	Modeler	Coupling Flow	Generation	Optimization
Rezende et al [83]	NF	Modeler	NF	Generation	Optimization
Gabrié et al. [151]	NF	Optimizer	NF	Generation	Optimization
Pearce et al. [152]	DM	Controller	DDPM	Imitation Learning	Robot Control /Games
Diffuser[153]	DM	Controller	DDPM	Offline RL	Structural Generation
Decision Diffuser [154]	DM	Controller	DDPM	Offline RL	Structural Generation
Decision Stacks [155]]	DM	Controller	DDPM	Offline RL	Structural Generation
Diffusion-QL [156]	DM	Controller	DDPM	Offline RL	Robot Control
Diffusion Policy [157]	DM	Controller	DDPM	Robotics	Robot Control
SfBC[158]	DM	Controller	DDPM	Offline RL	Structural Generation
UniPi[159]	DM	Controller	DDPM	Offline RL	Robot Control
AdaptDiffuser [160]	DM	Controller	DDPM	Robotics	Robot Control
DIPO[161]	DM	Controller	DDPM	Online RL	Robot Control
MTDiff [162]	DM	Modeler	DDPM	Offline RL	Robot Control
GenAug [163]	DM	Modeler	LDM	Robotics	Robot Control
SynthER [164]	DM	Modeler	EDM	Offline & Online RL	Robot Control
DDOM[165] Li et. al. [166]	DM	Optimizer	DDPM	Others	Optimization
DiffOPT[167]	DM	Optimizer	DDIM	Others	Optimization
	DM	Optimizer	DDIM	Others	Optimization
GFlowNets[51],[103] Pan et al. [168]	GFN GFN	Controller	GFN	Offline & Online RL	Structural Generation
GAFlowNets [169]	GFN	Controller Controller	GFN	Offline RL	Structural Generation
AFlowNets [170]	GFN		GFN	Offline RL	Structural Generation
Brunswic et al. [171]	GFN	Controller Controller	GFN	Offline RL	Structural Generation /Games
	CFN		GFN	Offline RL	Structural Generation
CFlowNets [104]		Controller	CFN	Online RL Generation	Structural Generation Optimization
Zhang et al.[172] Zhang et al. [173] GFACS[174]	GFN GFN GFN	Modeler Modeler	GFN GFN

TABLE 1: Generative models in decision making

Structure. The third dimension illustrates the basic model structure of generative models. Below is the structure for each category of generative models.

EBM: EBM

● GAN: GAN / WGAN / DCGAN / C-GAN.

● VAE: VAE / CVAE / seq2seq CVAE / β VAE / MAE.

NF: NF / Coupling Flow / Autoregressive Flow / Continuous Flow.

● DM: DDPM / DDIM / LDM / EDM.

● GFN: GFN / CFN.

● AM: Decoder Only / Encoder Decoder.

Expertise. The fourth dimension represents the expertise areas of the generative models in the sequential decision making. The expertise area includes Imitation Learning, online RL, offline RL, Robotics, Generation and Others.

Application. The last dimension highlights several representative applications where generative models have shown particular promise, including robot control, autonomous driving, gaming, structural generation, and optimization.

4 SURVEY: METHODOLOGY

This section delves into the three core functions of generative models and organizes the existing literature based on the methodology taxonomy we propose.

4.1 Generative Models as Controller

A controller aims to find a policy (a set of instructions or rules) that determines how to modify the system’s inputs to attain the desired behavior [186], [187]. The policy is typically generated through a process called control system design, which involves analyzing the system’s behavior, defining the desired outcomes, and then designing a controller to achieve those goals. The method used to generate the policy according to the system’s specific requirements and the design process’s objective. Generative models are able to model the decision-making process and produce diverse decision paths, making them well suited for generating policies in control systems.

Generative models are adept at capturing complex data distributions, making them particularly effective for learning from expert trajectories in decision-making tasks. When applied as controllers, these models can leverage offline data containing trajectories associated with high-reward values to extract and replicate optimal policies. In general, generative models facilitate the generation of control policies through two primary approaches, which we detail below.

(1) Generative models learn the policy from the stateaction or observation-action pairs, as shown in (19):

$$\tag{19}$$

where L is a metric function.

(2) Generative models learn the policy from the decision trajectories D = (s0, a0, s1, a1, ⋯) as shown in (20):

$$\tag{20}$$

The detailed explanations are as follows.

4.1.1 EBMs as Controller

EBMs serve as controllers by solving the Inverse Reinforcement Learning (IRL) problems, where the goal is to extract a control policy based on expert demonstrations. In this context, the energy-based model is used to learn the underlying reward structure of the expert’s behavior. Du & Mordatch [119] take a pioneering approach by using Markov Chain Monte Carlo (MCMC)-based EBM training to generate robotic hand trajectories, applying it to continuous neural networks. Building on this, EBIL [120] further refines the process by introducing a more straightforward solution to the IRL problems. Specifically, EBIL estimates the expert’s energy through score matching, which serves as a surrogate reward function. This reward function is then utilized to learn the policy through RL algorithms.

4.1.2 GANs as Controller

Generative Adversarial Imitation Learning (GAIL) [64] combines imitation learning with GANs to directly extract the policy from data. It seeks to find a saddle point (π, D), representing a pair consisting of a policy π and a discriminator D, by optimizing the following expression:

$$\tag{21}$$

where (s, a) denotes the state-action pair, π is the policy, πE represents the expert policy and H is a regularizer for the policy. To overcome some of GAIL’s limitations, several extensions have been proposed. For example, InfoGAIL [125], which builds upon GAIL, aims to calculate the potential structure of expert demonstrations in an unsupervised manner. This method can both imitate complex behaviors and extract meaningful, interpretable representations from intricate behavioral data. Similarly, MGAIL [126] incorporates a forward model into GAIL, making the calculation differentiable and enabling policy training with fewer samples and interactions with the environment. In contrast, Fail-Safe Adversarial Generative Imitation Learning (FAGIL) [127], an extension of GAIL, introduces a safety layer to ensure the generated policy is safe, while also facilitating closedform computation of probability densities and gradients. This method integrates end-to-end generative adversarial training with worst-case safety guarantees, making it wellsuited for scenarios requiring both safety and performance. On the other hand, WGAIL [128] and IC-GAIL [129] address the issue of imperfect demonstrations in GAIL. WGAIL focuses on learning the weights of imperfect demonstrations without requiring substantial prior information, while IC-GAIL uses confidence levels to assess the quality of these demonstrations, improving learning efficiency in the presence of noisy or suboptimal expert data.

Moreover, there are other related approaches such as Adversarial Inverse Reinforcement Learning (AIRL) [130], which formulates policy learning as an adversarial reward learning problem, allowing for acquisition of policies in diverse environments. AugAIRL [131] enhances AIRL by adopting semantic rewards in the learning framework (see Fig. 7 for an overview), while WAIL [132] bridges inverse reinforcement learning with optimal transport, using regularized optimal transport for large-scale applications. Lastly,

Model-based AIRL (MAIRL) [133] improves policy optimization by leveraging a self-attention dynamics model to make the computation graph end-to-end differentiable, and SC-AIRL [188] further reduces the need for extensive exploration by decomposing long-horizon tasks into subtasks, using shared rewards and critics across different subtasks.

Fig. 7: The augmented AIRL learning framework uses the discriminator’s weights as the reward function [131].

4.1.3 VAEs as Controller

Variational Autoencoders (VAEs) serve as controllers by learning compact, structured latent spaces that encode highdimensional environments and decision-making processes, which are directly used to generate control actions.

For example, Play-LMP [141] introduces a selfsupervised approach to learn control from play behaviors, enabling the system to autonomously generate control policies. Building on this idea, TACO-RL [142] extends Play-LMP by adopting a hierarchical model, which allows for learning long-horizon policies from high-dimensional camera observations in agnostic tasks. OPAL [143] utilizes VAE for offline learning of primitive actions in imitation learning tasks. Similarly, Generalization Through Imitation (GTI) [139] develops a two-stage algorithm that leverages VAE for policy learning through imitation, enhancing generalization in complex environments. Finally, HULC [140] attempts to improve performance by integrating hierarchical decomposition in robot-controlled learning, utilizing a multimodal transformer encoder, leveraging discrete latent plans, and applying a self-supervised contrastive loss.

Additionally, MaskDP [144] introduces a masked autoencoding technique that enhances decision-making models by learning latent spaces in a scalable and generalizable manner, thereby improving performance in highdimensional tasks (see Fig. 8). Similar to VAEs, this approach focuses on generating control policies by learning efficient latent representations, while additionally incorporating masking mechanisms to improve the model’s ability to generalize across tasks.

4.1.4 Normalizing Flows as Controller

Normalizing flows have garnered interest recently due to their capacity to learn complex probability distributions and perform efficient density estimation. While they are primarily used in generative modeling tasks, they can also be applied to decision-making problems control [75], [91]. In this context, normalizing flows can be leveraged to model and approximate the underlying distribution of the decision space [53], [89].

Fig. 8: Illustration of MaskDP. In the pretraining phase, the model is trained to predict masked tokens. After pretraining, it can be applied to a variety of downstream tasks by using different masking patterns [144].

The natural use of Normalizing flows for controllers is behavioral cloning, where the behavior to reproduce is represented by the target distribution µ. However, since the training only requires µ to be samplable and ν to have a defined density, one can train a ”reversed Normalization flows”, where ν represents a complex distribution defined by its density and µ is a simpler distribution, such as a normal distribution.

Normalizing Flow Policy (NF Policy) [148] integrates Normalizing Flow policies into the Soft Actor-Critic (SAC) framework, enabling the learning of more expressive policies. Fig. 9 illustrates the process of sampling actions from NF policies. Conservative Normalizing Flows (CNF) [149] makes use of Normalizing Flows as a conservative action encoder to learn policies in latent action spaces. First, the action encoder undergoes supervised pre-training on an offline dataset, and then a reinforcement learning-based policy model (serving as a controller in the latent space) is trained. Guided Flows [150] generates plan in the offline reinforcement learning setting by training Continuous Normalizing Flows based on regressing vector fields via integrating classifier free guidance.

Fig. 9: A sampled action from NF policies [148].

4.1.5 Diffusion Models as Controller

Diffusion models, characterized by their iterative refinement and controlled noise perturbation processes, have proven to be powerful frameworks for sequential decision-making tasks. These models excel in scenarios requiring the progressive transformation of noisy inputs into structured outputs, enabling them to serve as effective controllers.

There are generally two diffusion-based frameworks for decision making. The first approach utilizes diffusion models to generate subsequent actions based on historical observations, effectively capturing temporal dependencies in sequential tasks [189]. On the other hand, the second one employs diffusion models to learn the trajectory of the decision-making process, modeling the progression of decisions over time and capturing the broader decision path.

Diffusion models can abstract the distribution of observation-action pairs as a policy. For example, diffusion Q-learning (Diffusion-QL) [156] employs a conditional diffusion model as its policy. This model learns an action-value function and seeks to optimize actions that align closely with the behavior policy by optimizing the loss function. The policy is trained by reducing the loss below:

$$\tag{22}$$

SfBC [158] decomposes the policy into two components: an expressive generative behavior model for capturing the data distribution and an action evaluation model for assessing action quality. DIPO [161] utilizes a diffusion policy to perform model-free online reinforcement learning, enabling efficient decision-making in complex environments.

Diffusion models also provide an effective way to describe the joint distribution of trajectories. Diffuser [153] treats trajectory optimization as a modeling problem, which is solved by iteratively refining trajectories using a diffusion probabilistic model. In this framework, a trajectory is represented as τ = s0 s1 sT , where T denotes the time a 0 a1 aT horizon. During training, Diffuser minimizes a diffusion loss to optimize the trajectory generation process:

$$\tag{23}$$

Diffuser leverages classifier-guided sampling and image inpainting to reinterpret planning strategies. Decision Diffuser [154] models decision-making through conditional generative modeling, treating the policy as a returnconditioned diffusion model (see Fig. 10 for an overview). This approach allows Decision Diffuser to incorporate additional variables such as constraints and skills, enabling the generation of behaviors that meet multiple constraints simultaneously or demonstrate a combination of skills during test-time. AdaptDiffuser [160] creates rich synthetic expert data for goal-conditioned tasks by guiding the process with reward gradients and selecting high-quality data through a discriminator to finetune the diffusion model.

Diffusion models offer diverse applications in decisionmaking. Decision Stacks [155] deconstructs a goalconditioned policy agent into three generative modules that use independent models trained in parallel with teacher forcing to simulate the temporal evolution of observations, rewards, and actions. Diffusion Policy [157] simulates robotic policy by converting a visuomotor policy into a denoising diffusion process conditioned on specific inputs. UniPi [159] casts sequential decision-making as a video generation task conditioned on text. The planner generates a sequence of future frames depicting planned actions based on a text-encoded goal, from which control actions are derived. Pearce et al. [152] attempts to imitate human behavior using diffusion models, abstracting a rich distribution over the joint action space.

Fig. 10: Planning with decision diffuser via generative modeling and inverse dynamics [154].

4.1.6 GFlowNets as Controller

GFlowNet (GFN) is a powerful tool for sequential sampling, particularly in Markov decision processes (MDPs). In such a process, where a reward function R(⋅) is defined on each terminal state sf , GFlowNet can be trained to learn a stochastic policy that determines the probability of any terminal state sf is proportional to R(sf ). The most basic form of GFlowNet is applied to discrete, acyclic MDP (DAGs), with both the state and action space are finite, and the states and their transitions form a directed acyclic graph (DAG). In this context, we define four key components of a GFlowNet:

1) An edge flow function F (s, s′) which denotes the flow on the edge s → s′.

2) A vertex flow function F (s):

$$\tag{24}$$

3) Forward probability PF (s, s′) = F (s,s′)F (s) .

4) Backward probability PB(s, s′) = F (s′,s)F (s′) .

The induced policy of a GFlowNet is given by PF ,i.e., πf s,a s′ (s, a) = PF (s, s′).

GFlowNet [51], [103] is well-suited for complex scenarios where multiple trajectories may lead to the same terminal state. It transforms the set of trajectories into a flow and formulates the flow consistency equations as the learning objective, similar to how the Bellman equations are cast into Temporal Difference (TD) methods. GFlowNet demonstrates that any global minimum of the proposed objective results in a policy that samples from the desired distribution, achieving improved performance and diversity.

Fig. 11: Overall framework of CFlowNets [104]: environment interaction, flow estimation, and training.

Stochastic GFlowNet [168], an extension of GFlowNet, is tailored for stochastic environments. It decomposes state transitions into isolated environmental stochasticity, with a dynamics model trained to capture this stochasticity. GAFlowNets [169] introduces an intermediate reward as intrinsic motivation to address the exploration problem in sparse reward environments. This model combines edgebased and state-based intrinsic rewards to guide exploration effectively. Expected Flow Networks (EFlowNets) [170] adapts GFlowNets to stochastic environments as well, extending their applicability. Brunswic et al. [171] further generalizes GFlowNets to measurable spaces, including continuous state spaces, and extends the concept to handle cycles in this broader context.

CFlowNet [104] represents a significant advancement in the GFlowNet community by extending its capabilities to continuous control tasks (see Fig. 11). This extension is crucial as it bridges the gap between simplified simulation environments and complex real-world scenarios, expanding GFlowNet’s applicability to a broader range of practical applications. CFlowNet introduces a novel approach for solving continuous control tasks using GFlowNets. The authors present a theoretical formulation of CFlowNets and propose a unified framework that combines the action selection strategy, flow approximation method, and optimization objectives. Experimental results demonstrate that CFlowNets outperform many reinforcement learning methods, particularly in terms of exploration ability. A subsequent work, Continuous GFlowNet [106], develops a mathematical framework for continuous GFlowNet theory, further advancing the field.

4.1.7 Autoregressive Models as Controller

Autoregressive models excel in decision-making tasks by sequentially generating outputs conditioned on past observations. Decision Transformers (DTs) [176] advance this approach by defining reinforcement learning as conditional sequence modeling. It attempts to reformulate reinforcement learning as a conditional sequence modeling problem. It employs a causally masked Transformer to generate optimal actions aimed at maximizing the expected reward, conditioned on the past states, actions and the reward. The Online Decision Transformer [178] pretrains in the offline mode and fine-tunes online, unifying both within a framework that leverages sequence-level entropy regularizers and autoregressive modeling objectives to guide the learning process. DADT [190] improves the generalization ability by incorporating representation learning, which predicts the next state.

Fig. 12: Decision transformer architecture [176].

Prompt-DT [191] extends DTs by introducing taskspecific prompts as Fig. 13, enabling few-shot policy generalization and quick adaptation to new tasks, while maintaining the autoregressive decision-making framework.

Causal Decision Transformer for Recommender Systems (CDT4Rec) [192] applies DTs to recommendation systems by incorporating causal inference, predicting recommendations based on user history to maximize long-term reward. It serves as a controller, generating actions conditioned on past user interactions and system states.

Fig. 13: Prompt-DT for few-shot policy generalization [191].

Trajectory Transformer [177] also extends this perspective by framing RL as sequence modeling over full trajectories. It models distributions over trajectories using a Transformer architecture and employs beam search for planning, enabling the generation of high-reward action sequences. BooT [182] further refines this approach by leveraging bootstrapping techniques and generating additional selfsupervised offline data to enhance sequence model training.

Fig. 14: The trajectory transformer trains on autoregressively discretized state, action, and reward sequences, using a planning process akin to language model sampling [177].

Pareto-Efficient Decision Agents (PEDA) [181] extends Decision Transformers [176] and RvS [193] for multiobjective reinforcement learning. It introduces a preferenceand return-conditioned policy to handle multiple objectives efficiently in offline settings. With the advancement of largescale language models [2], Transformer-based controllers have emerged as versatile solutions capable of handling a diverse set of tasks within a single unified model. GATO [179] introduces a generalist agent capable of handling multi-modal, multi-task, and multi-embodiment scenarios. This agent is versatile in executing a wide range of functions, including chatting, controlling robotic arms, and determining outputs such as text, joint torques, button presses, or other actions based on contextual input. Based on this work, RT-1 [194] develops a scalable model trained on a large-scale dataset of real-world tasks performed by real robots. It shows performance improvements that scale with data size, model capacity, and the diversity of tasks involved.

4.2 Generative Models as Modeler

Unlike previous method S4RL [195] or RAD [196], which augments the states or the input spaces with slight perturbation to ensure the consistency, generative models take a different approach. Instead of merely modifying existing inputs, generative models focus on learning the underlying patterns within a dataset, thereby generating new data that closely resembles the original distribution [197].

In other words, a generative model processes a given set of input data and analyzes it to uncover the inherent patterns and structures of the data points [198]. In this context, generative models can be viewed as powerful ”modelers”, capable of capturing complex relationships within data and creating new and similar data that reflect these intrinsic relationships. This means that they can be employed to extract data representations for downstream tasks, such as classification or regression.

For example, generative models are capable of generating new images that resemble the images in the training dataset [199]. The model analyzes the features and patterns of the original images and uses this information to produce new images that mirror the style and content of the originals, just like DALL·E [1] does. Similarly, generative models can be used to create new text similar in style and content to the training corpus. Beyond image and text generation, they are useful for creating synthetic data for purposes like data augmentation, privacy protection, and simulations.

Overall, a generative model can be viewed as a ”modeler” because it learns the underlying patterns from a dataset [200] and generates new data that closely resembles the original data.

4.2.1 EBMs as Modeler

As ”models”, EBMs model data through an energy function that captures the underlying structure. Unlike other generative models, EBMs do not rely on explicit probabilities distributions but instead use an energy landscape to describe the data. Once trained, EBMs can generate new data by minimizing the energy, ensuring that the generated samples align with the learned patterns of the original dataset.

Boltzmann Machines (BM) [33] uses contrastive divergence techniques to generate samples that match the training data’s distribution. Similarly, DEBMs [121] extend this approach by using energy-based models to learn a distribution over state-action pairs, enabling reinforcement learning agents to generate and select optimal actions through contrastive divergence. Score-Based Generative Models (SGMs) [100] use score matching techniques to learn the gradients of the data distribution, enabling the generation of samples by refining noisy data towards the target distribution.

4.2.2 GANs as Modeler

GANs serve as models by learning the underlying patterns through their unique adversarial framework, which consists of two parts: the generator and the discriminator. Synthetic data is generated by the generator, with the discriminator assessing its realism, which in turn drives the generator to refine its output.

EGAN [134] serves as a modeler by leveraging the relationship between states and actions to learn the underlying patterns of the environment, which enables the pre-training of the agent and accelerates the learning process. Similarly, State2Pixel (S2P) [135] synthesizes raw pixel images from the agent’s state, effectively bridging the gap between the state space and the image domain in reinforcement learning (RL) algorithms. Additionally, S2P facilitates virtual exploration of the latent image distribution through model-based transitions in the state space.

4.2.3 VAEs as Modeler

VAEs are well-suited as modelers due to their capacity in learning a probabilistic mapping from the data space to the latent space, capturing the underlying variations in the data.

Han & Kim [145] represent the subspaces of the dataset by using the variational autoencoder. Subsequently, the VAE decoder is used to extract new data from the latent space and maps it back to the original one.

4.2.4 Normalizing Flows as Modeler

Normalizing Flows (NF) serve as powerful tools for modeling complex, high-dimensional data distributions by applying a series of invertible transformations. This flexibility enables NF to capture intricate data structures, making them particularly useful for tasks such as density estimation, variational inference, and posterior distribution modeling.

For instance, Non-linear Independent Component Estimation (NICE) [90] applied NF to model complex distributions by applying a series of invertible transformations, allowing it to represent high-dimensional data distributions flexibly. Rezende and Mohamed [83] further expanded the use of NF by applying them in variational inference to model complex posterior distributions. By applying NF to parameterize the posterior, they were able to construct flexible, tractable distributions that could be used for efficient Bayesian inference. This method helped overcome challenges associated with approximating posterior distributions, offering a scalable approach to variational inference in high-dimensional spaces. Muller et al. ¨ [201] enhanced NF’s flexibility by introducing piecewisepolynomial coupling transformations, which enhanced NF’s ability to model more complex distributions. By applying NF to parameterize the posterior, they were able to construct flexible, tractable distributions that could be used for efficient Bayesian inference. This method helped overcome challenges associated with approximating posterior distributions, offering a scalable approach to variational inference in high-dimensional spaces.

4.2.5 Diffusion Models as Modeler

Diffusion models are based on a network architecture that iteratively denoises samples, starting from random noise and gradually transforming it into structured data. This process enables them to capture complex data distributions and generate high-quality data by learning to reverse the noise-corruption process, making them highly effective as modelers.

Fig. 15: Overall architecture of M T DIF F . Different colors represent different tasks [162].

ROSIE [202] introduces text-to-image diffusion models to acquire valid data for robot to learn without additional robot data. Specifically, this method takes advantage of the art text-to-image diffusion. ROSIE inpaints diverse invisible objects for manipulation, backgrounds, and distractors with the guidance from the text. In robotics manipulation tasks like tabletop pick-and-place, GenAug [163] utilizes a small image-action demonstration dataset and produces an augmented dataset of image observations to enhance the real-world demonstration data. SynthER [164] flexibly upsamples an agent’s collected experience to train RL agents offline and online, in both proprioceptive and pixel-based environments. MTDiff [162] introduces a single diffusion model to establish large-scale multi-task offline data model, which combines Transformer backbones and prompt learning for data synthesis in multitask offline scenarios.

4.2.6 GFlowNets as Modeler

Generative Flow Networks (GFlowNets) model complex data distributions by learning to sample from rewarddefined distributions, making them effective for tasks like data augmentation and simulation.

In [172], GFlowNets treat sampling as a decision-making process, using Markovian trajectories to model data distributions. This unifies various generative models, enabling efficient training and inference algorithms.

[173] extends this idea with energy-based GFlowNets (EB-GFNs), combining GFlowNets with energy functions to generate samples from energy-defined distributions. This approach improves probabilistic modeling, particularly for high-dimensional discrete data.

In [203], GFlowNets are applied to text-to-image generation. The Diffusion Alignment with GFlowNet (DAG) algorithm post-trains diffusion models by generating highreward images with high probability, addressing the challenges of alignment in generative models.

4.2.7 Autoregressive Models as Modeler

Autoregressive models, such as PixelCNN [85] and PixelRNN [183], generate data by modeling the conditional dependencies between data points, predicting each component sequentially based on previous ones. This structure allows them to capture complex patterns in data, making them highly effective ”modelers”, especially in tasks like image generation, where they predict pixel distributions to generate coherent, high-quality images.

4.3 Generative Models as Optimizer

Recent research has increasingly approached optimization problems as sampling tasks, with generative models being used to tackle these challenges. This shift allows generative models to learn data representations that can be directly used to optimize specific objective functions, making them powerful tools for solving complex optimization problems [204]. These models can capture intricate data distributions and generate solutions that are both efficient and scalable [205]. In this context, we explore how generative models, when applied as optimizers, offer new perspectives on improving optimization strategies, surpassing traditional methods in several domains.

4.3.1 EBMs as Optimizer

Energy-Based Models (EBMs) can also function as powerful optimizers by refining candidate solutions toward optimal configurations in the solution space. The key mechanism involves learning an energy function that not only evaluates the quality of solutions but also guides the optimization process. EBMs optimize the energy landscape by employing techniques like contrastive divergence [206] or stochastic gradient descent. Rather than generating random samples, the optimization process is driven by minimizing the energy function, which allows EBMs to select the most optimal solutions according to the learned energy landscape. Methods like MCMC or Langevin dynamics are often employed to iteratively refine candidate solutions, effectively navigating the solution space to converge toward the global or local minimum.

EBMs have shown their effectiveness in various optimization tasks. One approach, SO-EBM [122], leverages EBMs to directly parameterize optimization problems through differentiable energy functions. Unlike traditional methods that rely on KKT conditions to implicitly induce an optimization layer, SO-EBM explicitly defines the optimization process by using an energy-based model, allowing for more accurate capture of the optimization landscape. This approach enables more efficient and refined optimization, particularly in complex, high-dimensional tasks.

Additionally, the Pareto-compositional energy-based model (pcEBM) [123] is designed for protein sequences optimization and sampling. In this approach, pcEBM integrates multiple objectives into a single energy function through a compositional structure that balances competing performance criteria. The model learns a multi-objective energy function that simultaneously accounts for factors such as structural stability and sequence diversity, which are critical in protein design. By optimizing this compositional energy function, pcEBM efficiently navigates the solution space to identify sequences that achieve a Paretooptimal balance. This method enhances the optimization process, making it particularly valuable for tasks requiring the simultaneous optimization of multiple criteria, such as in protein engineering and drug design. Similarly, a multistage coarse-to-fine expanding and sampling strategy (CF-EBM) [124] also utilizes sampling as a key mechanism for optimization. While pcEBM optimizes multi-objective tasks by refining protein sequences, the coarse-to-fine method enhances sample quality by progressively refining solutions. Both methods optimize the energy landscape through sampling, highlighting the versatility of EBMs in addressing a range of optimization challenges, from multi-objective tasks to high-quality generative sampling.

4.3.2 GANs as Optimizer

Generative Adversarial Networks (GANs) have emerged as powerful optimizers, addressing challenges in highdimensional and complex optimization problems.

He et al. [136] demonstrated how GANs can generate high-quality offspring solutions in multi-objective evolutionary algorithms, improving both solution diversity and optimality. This method alleviated the curse of dimensionality, improving optimization performance with limited data.

Similarly, Sim et al. [137] utilized GANs and DCGANs for topology optimization, generating valid data for structural design optimization. Their method utilized K-means clustering analysis to select optimized solutions, showing how GANs can be leveraged for structural design optimization, efficiently navigating complex design spaces and producing feasible solutions that meet specific criteria. This highlights GANs’ flexibility in optimization tasks beyond traditional settings, extending their application to engineering and design optimization.

Furthermore, Kalehbasti et al. [138] explored the combination of Conditional GANs (C-GANs) with genetic algorithms for solving high-dimensional nonlinear optimization problems. By augmenting solutions generated by a genetic algorithm with C-GANs, they enhanced the search for optimal solutions, demonstrating that the integration of GANs with evolutionary algorithms significantly accelerates the optimization process, even in the absence of vast amounts of training data. This development underscores the growing capability of GANs to serve as powerful optimizers in diverse optimization problems.

4.3.3 VAEs as Optimizer

VAEs are effective optimizers due to their ability to learn compact probabilistic representations, which facilitate efficient exploration of large and complex decision spaces. By mapping complex high-dimensional inputs to a simpler, lower-dimensional latent space, VAEs enable more efficient exploration of the decision space for optimization tasks.

For instance, CageBO [146] employs a conditional VAE to model the distribution of viable decisions in optimization tasks. By establishing a bidirectional mapping between the initial decision space and a reduced, unconstrained latent space, CageBO is able to optimize decision-making processes more efficiently. In particular, the model uses the latent representation to navigate the decision space, enabling more effective and scalable exploration of possible solutions. This approach improves optimization performance by generating high-quality candidate solutions and reducing the computational burden typically associated with exhaustive search methods.

Similarly, Conditional Variational Autoencoders (CVAEs) are also used in optimization tasks where the decision space is complex, such as routing problems. In [147], researchers employ a CVAE to learn a latent representation of the solution space, optimizing the search for feasible solutions (see Fig. 16). By conditioning the generative process on problem-specific constraints, CVAE refines the optimization process to generate high-quality solutions within the reduced latent space.

Fig. 16: The overall training process of CVAE-Opt (Left) and the iterative search process (Right) [147].

4.3.4 Normalizing Flows as Optimizer

Normalizing Flows (NF) can also serve as powerful optimizers. The core idea behind using Normalizing Flows (NF) as an optimizer is their ability to transform simple distributions (like Gaussian or uniform) into complex ones that match the desired solution distribution. This enables the generation of high-quality solutions efficiently, with fewer iterations than traditional methods. Because NF are differentiable, they can be easily integrated with gradient-based optimization, allowing for smooth, end-to-end optimization through backpropagation.

Gabrie et al. [151] proposed an adaptive MCMC method ´ that enhances traditional MCMC sampling by incorporating nonlocal transition kernels parameterized by normalized flows. This method enhances optimization by using NF to guide the search more efficiently. By combining local MCMC updates with NF-driven transitions, it accelerates convergence to optimal solutions, improving both efficiency and performance.

4.3.5 Diffusion Models as Optimizer

Diffusion models have shown increasing potential as optimizers. The core advantage of diffusion models lies in their iterative denoising process, which allows them to refine candidate solutions over time, making them particularly effective for black-box optimization tasks where the objective function is unknown or difficult to compute directly.

In traditional optimization problems, gradient-based methods often struggle with high-dimensional spaces, especially when the objective function is highly complex or nonconvex. In contrast, diffusion models approach optimization through a process of gradually refining solutions, effectively exploring the solution space and generating promising candidates.

For instance, Denoising Diffusion Optimization Models (DDOM) [165] introduce a novel inverse approach for solving offline black-box optimization problems by leveraging diffusion models. This method models complex objective functions through a reverse diffusion process, offering an efficient way to explore and optimize in high-dimensional, poorly understood spaces. By iterating through multiple diffusion steps, DDOM can refine candidate solutions to find near-optimal results even when the objective function is opaque or difficult to compute directly.

Similarly, Li et al. [166] reformulate optimization as a conditional sampling task and propose a reward-directed conditional diffusion model as Fig. 17. This model is trained using mixed data to sample near-optimal solutions conditioned on predicted rewards, guiding the optimization process in a direction that maximizes the objective function’s expected value. This reward-directed approach allows the model to focus on the most promising areas of the solution space, further enhancing its efficiency in optimization tasks.

Fig. 17: Black-box optimization with reward-directed conditional diffusion models [166].

Another approach, DiffOPT [167], reframes the optimization problem by drawing samples from a joint Boltzmann distribution determined by both the objective function and the learned data distribution. This method incorporates a two-stage framework: the first stage uses a guided diffusion process to initialize the optimization, while the second stage applies Langevin dynamics to further refine the solution. The combination of these two stages allows for both efficient exploration and precise refinement, making DiffOPT particularly effective for complex optimization tasks.

4.3.6 GFlowNets as Optimizer

Generative Flow Networks (GFlowNets) have emerged as powerful optimizers, effectively navigating complex solution spaces through a unique generative flow framework. By constructing solutions step by step according to a learned flow, GFlowNets can efficiently explore and exploit the underlying structure of optimization problems. The flowbased nature of GFlowNets enables them to generate diverse candidate solutions while maintaining high sample efficiency, which is crucial in scenarios where the evaluation of solutions is computationally expensive. Their ability to learn from both on-policy and off-policy data further enhances their adaptability and performance across a wide range of optimization tasks.

[207] explores the application of GFlowNets to combinatorial optimization (CO) problems, which are often NP-hard and challenging for exact algorithms. The researchers design a MDP tailored to various combinatorial problems and train conditional GFlowNets to sample from the solution space. Extensive experiments on diverse CO tasks with both synthetic and realistic data have demonstrated that GFlowNet policies can effectively find high-quality solutions.

Building on this foundation, GFlowNets were extended to multi-objective optimization, addressing the challenge of balancing multiple conflicting objectives. MOGFNs [175], a novel method based on GFLowNets, can generate diverse Pareto-optimal solutions by learning to navigate the tradeoffs between different objectives. This development significantly broadened the scope of GFlowNets’ applicability in optimization.

Additionally, Generative Flow Ant Colony Sampler (GFACS) [174] hierarchically combines amortized inference and parallel stochastic search. The solution sampling process is shown in Fig. ??. This hybrid approach leverages the strengths of both methods: the efficient sampling and flowbased learning of GFlowNets, and the heuristic search capabilities of ACO. GFACS demonstrated significant improvements in solving combinatorial optimization problems by incorporating off-policy training and local search strategies.

Fig. 18: Solution sampling mechanism of GFACS [166]. The GNN, trained with a GFlowNet loss, acts as an expert guide to select the next step in constructing a solution, like a tour in the TSP.

4.3.7 Autoregressive Models as Optimizer

Autoregressive models, particularly due to their sequential structure, are effective optimizers because they capture temporal or sequential dependencies between data points, allowing for iterative refinement of solutions. This structure enables the models to handle uncertainty and adaptively improve the optimization performance across various tasks. The sequential nature of autoregressive models, which predicts the next step based on previous steps, allows them to generate or adjust solutions gradually, making them highly effective in settings where the solution space is complex and high-dimensional.

For instance, Transformer Neural Processes (TNPs) [185] introduce a meta-learning framework that employs a transformer-based autoregressive model for optimizing performance under uncertainty in tasks that involve variable inputs or distributions. By leveraging the strengths of transformers in sequence modeling and combining them with Neural Processes for better uncertainty estimation, TNPs optimize a model’s performance on various tasks by iteratively refining predictions and learning optimal strategies.

Similarly, BONET [184] proposes a generative framework that uses autoregressive models for pretraining optimization on offline datasets. By training a transformerbased model in an unsupervised manner, BONET learns to optimize task performance in real-world settings without direct task-specific supervision. This enables it to adapt to a wide variety of optimization problems, from machine learning to more general generative optimization tasks.

Fig. 19: Illustration of the TNP-A architecture [185].

5 APPLICATIONS

There are lots of applications of generative models in decision making scenarios. We consider five typical applications including, robot control, autonomous driving, games, structural generation, and optimization.

5.1 Robot Control

Robot control refers to the process of commanding a robot to execute specific tasks or actions [208], [209], [210], [211], [212]. This can be achieved through a variety of methods, including manual control by a human operator, preprogrammed instructions, or autonomous decision-making using sensors and machine learning algorithms.

Policy generation in robotics refers to creating the decision-making framework that guides a robot’s actions [213], [214], [215]. It entails defining objectives or goals for the robot while incorporating constraints and limitations to ensure safe and efficient behavior.

Effective policy generation for robots requires consideration of various factors, such as the robot’s capabilities, limitations, operating environment, and task requirements. This often involves designing sophisticated algorithms and decision-making models capable of real-time adaptation to dynamic conditions.

Robot control generally encompasses several key aspects, including robot manipulation [216], trajectory generation [55], [107], [217], and locomotion [218]. Generative models play a pivotal role in either directly controlling robots or generating synthetic data to enhance the training of more effective control policies.

5.2 Structural Generation

Generative models are increasingly applied to various graph-based tasks, such as graph generation [219], graph completion [220], and graph classification [221]. In graph generation, these models learn the underlying structures of training graphs and use this knowledge to generate new graphs with similar characteristics. This capability has wide-ranging applications, including molecule design in chemistry, protein interaction modeling in biology, and architecture optimization in engineering.

A notable example of applying generative models in structure generation is through GFlowNets. Jain et al. [222] proposed an active learning framework that integrates epistemic uncertainty estimation with GFlowNets to generate diverse and informative candidate solutions for drug discovery and other problems. This method selects a batch of promising candidates after each learning round. To improve the exploration efficiency, Kim et al. [223] introduced a local search strategy within GFlowNets, targeting highreward sample spaces while preventing over-exploration. This approach combines forward and backward policies with backtracking and reconstruction techniques, further refining the generated solutions.

Recent works [224], [225], [226] have made notable advancements in integrating reinforcement learning (RL) with natural language processing (NLP). Xin et al. [224] propose Prompt-Based Reinforcement Learning (PRL), which uses historical data and state-reward pairs as prompts to train RL-based recommendation agents. Deng et al. [225] introduce RLPROMPT, an efficient method for discrete prompt optimization that leverages RL to discover optimal prompts for pre-trained language models (LMs). This approach is applicable across different types of LMs and tasks, including both classification and generation. Zhang et al. [226] present TEMPERA, which designs optimal prompts for large LMs in zero-shot or few-shot learning settings by leveraging prior knowledge and enabling flexible prompt edits.

These works illustrate how RL-optimized generative approaches, like prompt-based methods, improve decisionmaking in language models. By structuring inputs and outputs flexibly, they facilitate context-aware solutions for complex tasks, such as multi-step reasoning and dynamic adaptation, with potential applications beyond NLP in fields like robotics and planning.

5.3 Games

Games artificial intelligence (AI) is a prominent research area focused on developing AI techniques that achieve human-level performance in gameplay [227]. Games offer intriguing and complex problems for AI agents to solve within safe and controllable environments. They also provide diverse challenges, from strategic planning to realtime decision-making, serving as benchmarks for testing and advancing AI algorithms.

Several generative models have been developed for gameplay. Adversarial Flow Networks (AFlowNets) [170] extend EFlowNets to adversarial environments, specifically for two-player zero-sum games. GATO [179] is a versatile generalist agent capable of playing multiple games. Similarly, the Multi-Game Decision Transformer [180] adapts a single transformer-based model to handle diverse gaming scenarios.

5.4 Autonomous Driving

Generative models have been used in autonomous driving for a variety of tasks, including driving control [131], [228], [229], image and video processing [229], object detection [230] and scene understanding [231].

Huang et al. [232] introduced a hybrid framework that integrates neural decision-making into the classical modular pipeline using end-to-end imitation learning via reparameterized generative adversarial learning. This approach retains the advantages of the classical pipeline, such as strict adherence to physical and logical constraints, while enabling the framework to learn complex driving decisions directly from data.

Another application of generative model in autonomous driving is in the generation of synthetic data for training deep learning models [233]. Synthetic data helps augment real-world datasets, which are often constrained in size and diversity. These models can produce synthetic images and videos resembling real-world scenarios, thereby enhancing the performance of deep learning models in autonomous driving tasks. For example, TrafficGen [234] acts as an autoregressive generative model based on an encoder-decoder structure. It is trained on fragmented human driving data gathered from real-world environments to create lifelike traffic scenarios. By employing an attention mechanism, TrafficGen encodes the current traffic context. subsequently, it decodes the initial state of the vehicle and generates a long-term trajectory.

Generative models are valuable for processing sensor data, such as LiDAR and camera inputs, in applications like object detection and tracking [235]. These models can generate 3D representations of the environment from sensor data, which serve as a foundation for more effective object detection and tracking.

Generative model can also be used for scene understanding in autonomous driving [96]. For example, it can be used to generate semantic maps of the environment, which can be used for navigation and path planning. Generative model can also be used to generate predictions of future vehicle and pedestrian movements, which can be used for collision avoidance and planning.

PlanCP [236] leverages Conformal Prediction (CP) to quantify the uncertainty in diffusion dynamics models. By utilizing a finite set of exchangeable expert trajectory examples, PlanCP enhances robustness across diverse training scenarios, ensuring more reliable performance in autonomous driving applications.

Fig. 20: The bar chart illustrates the continuous growth of generative models in decision-making across various applications, as indicated in the legend. The methods highlighted in the figure represent the seminal works of each period.

Overall, generative models hold significant potential for enhancing the performance and safety of autonomous driving systems [237]. They are versatile tools applicable to a range of tasks, including data augmentation and scene understanding, and have shown promising results in both research studies and practical applications.

5.5 Optimization

We consider three optimization settings: black-box optimization, neural architecture search, and scheduling. These settings represent critical challenges in decision-making tasks, where generative models have demonstrated significant potential.

In the black-box optimization setting, an agent learns to optimize an unknown function through pointwise evaluations [238]. For example, Transformer Neural Process [185] reframes meta learning with uncertainty awareness in the context of sequence modeling, solving it with an autoregressive likelihood-based objective, utilizing a novel transformer-based architecture.

Generative models such as GFlowNets have also been applied to combinatorial optimization. Zhang et al. [239] applied Markov decision processes (MDPs) to various combinatorial problems, training conditional GFlowNets to sample from solution spaces. DIFUSCO [240] introduced a graph-based diffusion framework to solve NP-complete problems by casting them as binary vector optimization tasks and leveraging graph-based denoising diffusion models to create superior methods.

Neural Combinatorial Optimization (NCO) approaches focus on leveraging neural networks to learn feasible solution distributions for combinatorial optimization problems [241], [242]. Autoregressive construction solvers [147], [243], [244], [245], [246] are built upon the success of transformerbased [109] architectures in sequential generation tasks. Non-autoregressive construction solvers [247], [248], [249], [250], [251], [252] have also been proposed to learn high quality solution distributions.

Furthermore, prompt tuning and reinforcement learning (RL)-based prompt optimization methods are practical applications of black-box optimization. Prompt tuning [253], [254], [255], [256] optimizes task-specific embeddings to guide model behavior efficiently, while RL-based methods like RLPROMPT [225] and TEMPERA [226] refine prompt design to improve performance in tasks like classification and few-shot learning. These approaches demonstrate how generative models can navigate vast solution spaces effectively, balancing exploration and exploitation.

In the neural architecture search (NAS) setting. NAS focuses on optimizing neural network designs to maximize task performance. GFlowOut [257] leverages GFlowNets to learn the posterior distribution of dropout masks, providing an innovative method for optimizing network architecture. By utilizing generative models, GFlowOut highlights the potential for improving architecture efficiency while maintaining high performance.

In the scheduling setting, optimization involves designing strategies to allocate resources or order tasks efficiently. GFlowNets have been deployed in compiler scheduling [258], where they control the balance between the diversity and quality of the proposed methods. This application underscores the ability of generative models to manage competing objectives in complex, high-stakes environments, such as industrial and computational settings.

These three optimization settings illustrate the versatility of generative models in tackling diverse and complex challenges, offering promising avenues for future exploration and development.

6 PERSPECTIVE: NEXT GENERATION GENERA-TIVE DECISION MAKING MODELS

Despite numerous successful examples, advancing nextgeneration generative decision-making models remains a critical challenge. In this section, we highlight three pivotal perspectives to enhance the capabilities of generative models: high-performance algorithms, large-scale generalized decision-making Models, and self-evolving and adaptive generative models.

6.1 High performance Algorithms

The growing integration of generative models in decisionmaking highlights an increasing demand for highperformance algorithms capable of addressing the complexities of real-world applications. Future research should focus on enhancing their scalability, efficiency, and adaptability to dynamic environments. Normalizing flows and GFlowNets represent two promising directions for advancing generative decision-making algorithms, each with unique strengths and challenges.

Normalizing flows have demonstrated potential in modeling uncertainties and supporting principled decisionmaking in complex, high-dimensional environments. However, several critical challenges remain. These include designing more effective objective functions tailored to decision-making tasks, improving their efficiency in handling high-dimensional decision spaces, and reducing the computational costs of training. Future work should explore novel architectures and optimization strategies to address these limitations, enabling their broader application in dynamic, real-world scenarios.

Similarly, GFlowNets present a novel probabilistic approach to modeling unnormalized probability distributions, making them particularly effective for generating structured or discrete objects such as graphs. Unlike VAEs and GANs, whose goals are to approximate data distributions or generate realistic samples, GFlowNets are specifically designed to sample x in proportion to a predefined reward function R(x). The objective ensures that the generated samples approximately follow the desired reward distribution.

By leveraging their strengths and addressing current limitations, these frameworks have the potential to reshape generative decision-making, enabling more efficient, scalable, and adaptable systems for complex, dynamic realworld applications.

6.2 Large-scale Generalized Decision Making Models

As artificial intelligence continues to advance towards large-scale models, developing generative large models for decision-making reasoning becomes increasingly essential. These models have the potential to handle complex reasoning tasks, integrate diverse knowledge, and offer robust solutions in dynamic and uncertain environments.

Additionally, current generative models are often tailored to predefined tasks or scenarios, limiting their generalization capabilities. To address real-world challenges, there is a pressing need for large-scale, generalized generative decision models capable of adapting to diverse and dynamic environments.

For instance, web-scale diffusion models represent a significant step toward achieving this generalization. These models leverage massive datasets and scalable architectures to learn intricate patterns across diverse domains, making them suitable for generating solutions in complex decisionmaking tasks. Similarly, large language models (LLMs), such as GPT [20] or PaLM [259], showcase the ability to process diverse inputs and perform multi-step reasoning, highlighting their potential as general-purpose decision-making tools. Moreover, vision-language-action (VLA) models push this generalization further by integrating multimodal data, enabling them to interpret, reason, and act across diverse scenarios. These examples underline the transformative potential of generative models in building adaptive, large-scale decision-making frameworks.

Besides, generative models should be capable of producing diverse solutions, enabling them to navigate various possibilities and adapt to the complexities in real-world decision-making scenarios. This diversity is particularly important in handling ambiguous or multi-objective tasks, where a single solution may not suffice. By generating varied yet high-quality outputs, these models can better account for uncertainties, provide alternative strategies, and support robust decision-making processes. Such capabilities are critical for applications in dynamic and multi-faceted domains, ranging from autonomous systems to large-scale optimization tasks.

6.3 Self-Evolving and Adaptive Generative Models

The third perspective emphasizes the importance of selfevolving capabilities in generative models to address dynamic and complex decision-making environments. While current generative models are often built with static assumptions and trained on fixed datasets, real-world decision-making scenarios require models that can adapt continuously, learn in real time, and transfer knowledge across different domains.

A key component of adaptive generative models is continual learning, which allows them to evolve as new data becomes available. This ability allows them to update their knowledge without forgetting previously learned information, a challenge known as catastrophic forgetting. By leveraging incremental training, memory replay, and regularization techniques, these models can respond to evolving data distributions and changing environments, ensuring their decision-making remains accurate and relevant. This adaptability is crucial for applications where data patterns shift frequently and models need to stay up-to-date without retraining from scratch.

Furthermore, the integration of real-time feedback mechanisms allows generative models to quickly adjust their decision-making strategies based on immediate outcomes or user interactions. Real-time performance monitoring can trigger dynamic updates, enabling the model to refine its approach as it receives feedback from the environment. This is essential for tasks that require quick reactions, such as autonomous navigation or adaptive recommendation systems, where rapid changes in input or context need to be addressed effectively to maintain high performance.

Cross-domain adaptability is another vital aspect, as real-world decision-making often spans multiple domains with distinct challenges and data characteristics. Generative models that can transfer knowledge learned from one domain to another can enhance their versatility and expand their applicability. Approaches like meta-learning and transfer learning enable models to recognize and apply shared structures and features across different tasks, making them capable of tackling new and unseen problems with minimal additional training. This cross-domain capability ensures that models are not limited to a narrow set of conditions but are instead able to operate effectively in diverse and multi-faceted environments.

Incorporating adaptive features like continual learning, real-time feedback, and cross-domain adaptability allows generative models to better align with real-world challenges. This approach paves the way for more robust, versatile, and future-proof decision-making systems that can thrive in complex and ever-changing environments, extending the potential of generative models far beyond their current scope and applications.

7 CONCLUSION

This survey systematically examines the applications of generative models in decision-making tasks. We begin by revisiting the problem formulation and comparing traditional and generative model approaches. We then present a taxonomy classifying generative models into seven categories based on their model families, functionalities, structures, and advantages. Next, we explore their core principles and real-world applications. We also outline promising future directions, emphasizing the desired characteristics of next-generation decision-making models. Leveraging their ability to model complex data distributions and generate innovative strategies, we argue that generative models hold transformative potential in decision-making. By synthesizing recent advancements and highlighting key research trajectories, this survey aims to bridge gaps and lay the groundwork for the continued evolution of generative models in decision-making.

REFERENCES

[1] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,” in International conference on machine learning. Pmlr, 2021, pp. 8821–8831.

[2] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020.

[3] J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo et al., “Improving image generation with better captions,” Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, vol. 2, no. 3, p. 8, 2023.

[4] R. OpenAI, “Gpt-4 technical report. arxiv 2303.08774,” View in Article, vol. 2, no. 5, 2023.

[5] R. Bellman, “A markovian decision process,” Journal of mathematics and mechanics, pp. 679–684, 1957.

[6] V. Krishnamurthy, Partially observed Markov decision processes. Cambridge university press, 2016.

[7] M. J. Kochenderfer, T. A. Wheeler, and K. H. Wray, Algorithms for decision making. MIT press, 2022.

[8] R. Hooke and T. A. Jeeves, ““direct search”solution of numerical and statistical problems,” Journal of the ACM (JACM), vol. 8, no. 2, pp. 212–229, 1961.

[9] S. Kirkpatrick, C. D. Gelatt Jr, and M. P. Vecchi, “Optimization by simulated annealing,” science, vol. 220, no. 4598, pp. 671–680, 1983.

[10] F. Tao, “Data-driven decision-making based on noisy data samples—studies in the machine learning applications,” Ph.D. dissertation, The University of Texas at San Antonio, 2021.

[11] R. S. Sutton, A. G. Barto et al., Introduction to reinforcement learning. MIT press Cambridge, 1998, vol. 135.

[12] D. Ernst, P. Geurts, and L. Wehenkel, “Tree-based batch mode reinforcement learning,” Journal of Machine Learning Research, vol. 6, 2005.

[13] J. Fu, A. Kumar, M. Soh, and S. Levine, “Diagnosing bottlenecks in deep q-learning algorithms,” in International Conference on Machine Learning. PMLR, 2019, pp. 2021–2030.

[14] R. S. Sutton, D. McAllester, S. Singh, and Y. Mansour, “Policy gradient methods for reinforcement learning with function approximation,” Advances in neural information processing systems, vol. 12, 1999.

[15] T. Degris, P. M. Pilarski, and R. S. Sutton, “Model-free reinforcement learning with continuous action in practice,” in 2012 American Control Conference (ACC). IEEE, 2012, pp. 2177–2182.

[16] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, “Trust region policy optimization,” in International conference on machine learning. PMLR, 2015, pp. 1889–1897.

[17] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017.

[18] J. G. Kuba, R. Chen, M. Wen, Y. Wen, F. Sun, J. Wang, and Y. Yang, “Trust region policy optimisation in multi-agent reinforcement learning,” arXiv preprint arXiv:2109.11251, 2021.

[19] C. Yu, A. Velu, E. Vinitsky, J. Gao, Y. Wang, A. Bayen, and Y. Wu, “The surprising effectiveness of ppo in cooperative multi-agent games,” Advances in Neural Information Processing Systems, vol. 35, pp. 24 611–24 624, 2022.

[20] A. Radford, K. Narasimhan, T. Salimans, I. Sutskever et al., “Improving language understanding by generative pre-training,” 2018.

[21] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,” arXiv preprint arXiv:1509.02971, 2015.

[22] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor,” in International conference on machine learning. PMLR, 2018, pp. 1861–1870.

[23] K. Chua, R. Calandra, R. McAllister, and S. Levine, “Deep reinforcement learning in a handful of trials using probabilistic dynamics models,” Advances in neural information processing systems, vol. 31, 2018.

[24] K. Asadi, D. Misra, and M. Littman, “Lipschitz continuity in model-based reinforcement learning,” in International Conference on Machine Learning. PMLR, 2018, pp. 264–273.

[25] T. Xu, Z. Li, and Y. Yu, “Error bounds of imitating policies and environments,” Advances in Neural Information Processing Systems, vol. 33, pp. 15 737–15 749, 2020.

[26] W. Zhang, Z. Yang, J. Shen, M. Liu, Y. Huang, X. Zhang, R. Tang, and Z. Li, “Learning to build high-fidelity and robust environment models,” in Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13–17, 2021, Proceedings, Part I 21. Springer, 2021, pp. 104–121.

[27] C. Sutton, A. McCallum et al., “An introduction to conditional random fields,” Foundations and Trends® in Machine Learning, vol. 4, no. 4, pp. 267–373, 2012.

[28] V. Feinberg, A. Wan, I. Stoica, M. I. Jordan, J. E. Gonzalez, and S. Levine, “Model-based value estimation for efficient model-free reinforcement learning,” arXiv preprint arXiv:1803.00101, 2018.

[29] R. Gozalo-Brizuela and E. C. Garrido-Merchan, “A survey of ´ generative ai applications,” arXiv preprint arXiv:2306.02781, 2023.

[30] R. Gozalo-Brizuela and E. C. Garrido-Merchan, “Chatgpt is not all you need. a state of the art review of large generative ai models,” arXiv preprint arXiv:2301.04655, 2023.

[31] A. Jo, “The promise and peril of generative ai,” Nature, vol. 614, no. 1, pp. 214–216, 2023.

[32] E. Brynjolfsson, D. Li, and L. R. Raymond, “Generative ai at work,” National Bureau of Economic Research, Tech. Rep., 2023.

[33] D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, “A learning algorithm for boltzmann machines,” Cognitive science, vol. 9, no. 1, pp. 147–169, 1985.

[34] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020.

[35] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-toimage translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2223–2232.

[36] J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International Conference on Machine Learning. PMLR, 2015, pp. 2256–2265.

[37] A. Verine, “Quality and diversity in generative models through the lens of f-divergences,” Ph.D. dissertation, Universite Paris ´ sciences et lettres, 2024.

[38] M. L. Li and S. Zhu, “Balancing optimality and diversity: Humancentered decision making through generative curation,” arXiv preprint arXiv:2409.11535, 2024.

[39] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020.

[40] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,” Advances in neural information processing systems, vol. 29, 2016.

[41] D. P. Kingma and P. Dhariwal, “Glow: Generative flow with invertible 1x1 convolutions,” Advances in neural information processing systems, vol. 31, 2018.

[42] A. A. Alemi, I. Fischer, J. V. Dillon, and K. Murphy, “Deep variational information bottleneck,” arXiv preprint arXiv:1612.00410, 2016.

[43] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013.

[44] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” arXiv preprint arXiv:1511.06434, 2015.

[45] A. Oussidi and A. Elhassouny, “Deep generative models: Survey,” in 2018 International conference on intelligent systems and computer vision (ISCV). IEEE, 2018, pp. 1–8.

[46] P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021.

[47] T. Salimans, A. Karpathy, X. Chen, and D. P. Kingma, “Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications,” arXiv preprint arXiv:1701.05517, 2017.

[48] L. Girin, S. Leglaive, X. Bie, J. Diard, T. Hueber, and X. Alameda-Pineda, “Dynamical variational autoencoders: A comprehensive review,” arXiv preprint arXiv:2008.12595, 2020.

[49] C. Finn, P. Christiano, P. Abbeel, and S. Levine, “A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models,” arXiv preprint arXiv:1611.03852, 2016.

[50] M. Abdollahzadeh, T. Malekzadeh, C. T. Teo, K. Chandrasegaran, G. Liu, and N.-M. Cheung, “A survey on generative modeling with limited data, few shots, and zero shot,” arXiv preprint arXiv:2307.14397, 2023.

[51] Y. Bengio, S. Lahlou, T. Deleu, E. J. Hu, M. Tiwari, and E. Bengio, “Gflownet foundations,” arXiv preprint arXiv:2111.09266, 2021.

[52] E. Betzalel, C. Penso, A. Navon, and E. Fetaya, “A study on the evaluation of generative models,” arXiv preprint arXiv:2206.10935, 2022.

[53] G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed, and B. Lakshminarayanan, “Normalizing flows for probabilistic modeling and inference,” The Journal of Machine Learning Research, vol. 22, no. 1, pp. 2617–2680, 2021.

[54] D. Belanger, “Deep energy-based models for structured prediction,” 2017.

[55] Y. Xu, T. Zhao, C. Baker, Y. Zhao, and Y. N. Wu, “Learning trajectory prediction with continuous inverse optimal control via langevin sampling of energy-based models,” arXiv preprint arXiv:1904.05453, 2019.

[56] J. J. Hopfield, “Hopfield network,” Scholarpedia, vol. 2, no. 5, p. 1977, 2007.

[57] M. Suhail, A. Mittal, B. Siddiquie, C. Broaddus, J. Eledath, G. Medioni, and L. Sigal, “Energy-based learning for scene graph generation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 13 936–13 945.

[58] R. B. Lyngsø and C. N. Pedersen, “Rna pseudoknot prediction in energy-based models,” Journal of computational biology, vol. 7, no. 3-4, pp. 409–427, 2000.

[59] J. Ngiam, Z. Chen, P. W. Koh, and A. Y. Ng, “Learning deep energy models,” in Proceedings of the 28th international conference on machine learning (ICML-11), 2011, pp. 1105–1112.

[60] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Distributed representations of words and phrases and their compositionality,” Advances in neural information processing systems, vol. 26, 2013.

[61] G. Parisi, “Correlation functions and computer simulations,” Nuclear Physics B, vol. 180, no. 3, pp. 378–384, 1981.

[62] A. Hyvarinen and P. Dayan, “Estimation of non-normalized ¨ statistical models by score matching.” Journal of Machine Learning Research, vol. 6, no. 4, 2005.

[63] M. Gutmann and A. Hyvarinen, “Noise-contrastive estimation:¨ A new estimation principle for unnormalized statistical models,” in Proceedings of the thirteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2010, pp. 297–304.

[64] J. Ho and S. Ermon, “Generative adversarial imitation learning,” Advances in neural information processing systems, vol. 29, 2016.

[65] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” in International conference on machine learning. PMLR, 2017, pp. 214–223.

[66] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Courville, “Improved training of wasserstein gans,” Advances in neural information processing systems, vol. 30, 2017.

[67] T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of gans for improved quality, stability, and variation,” arXiv preprint arXiv:1710.10196, 2017.

[68] C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint arXiv:1606.05908, 2016.

[69] T. D. Kulkarni, W. F. Whitney, P. Kohli, and J. Tenenbaum, “Deep convolutional inverse graphics network,” Advances in neural information processing systems, vol. 28, 2015.

[70] G. Liu, F. A. Reda, K. J. Shih, T.-C. Wang, A. Tao, and B. Catanzaro, “Image inpainting for irregular holes using partial convolutions,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 85–100.

[71] A. Deshpande, J. Lu, M.-C. Yeh, M. Jin Chong, and D. Forsyth, “Learning diverse image colorization,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 6837–6845.

[72] I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” in International conference on learning representations, 2016.

[73] J. Hu, J. Whitman, and H. Choset, “Glso: Grammar-guided latent space optimization for sample-efficient robot design automation,” in Conference on Robot Learning. PMLR, 2023, pp. 1321– 1331.

[74] J. An and S. Cho, “Variational autoencoder based anomaly detection using reconstruction probability,” Special lecture on IE, vol. 2, no. 1, pp. 1–18, 2015.

[75] K. Sohn, H. Lee, and X. Yan, “Learning structured output representation using deep conditional generative models,” Advances in neural information processing systems, vol. 28, 2015.

[76] N. Dilokthanakul, P. A. Mediano, M. Garnelo, M. C. Lee, H. Salimbeni, K. Arulkumaran, and M. Shanahan, “Deep unsupervised clustering with gaussian mixture variational autoencoders,” arXiv preprint arXiv:1611.02648, 2016.

[77] L. Mescheder, S. Nowozin, and A. Geiger, “Adversarial variational bayes: Unifying variational autoencoders and generative adversarial networks,” in International conference on machine learning. PMLR, 2017, pp. 2391–2400.

[78] J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in International Conference on Machine Learning. PMLR, 2021, pp. 5530– 5540.

[79] G. Silvestri, D. Roos, and L. Ambrogioni, “Deterministic training of generative autoencoders using invertible layers,” in The Eleventh International Conference on Learning Representations, 2022.

[80] F. Ye and A. G. Bors, “Continual variational autoencoder learning via online cooperative memorization,” in European Conference on Computer Vision. Springer, 2022, pp. 531–549.

[81] Z. Li, X. Jiang, R. Missel, P. K. Gyawali, N. Kumar, and L. Wang, “Continual unsupervised disentangling of self-organizing representations,” in The Eleventh International Conference on Learning Representations, 2022.

[82] J. Bae, M. R. Zhang, M. Ruan, E. Wang, S. Hasegawa, J. Ba, and R. Grosse, “Multi-rate vae: Train once, get the full rate-distortion curve,” arXiv preprint arXiv:2212.03905, 2022.

[83] D. Rezende and S. Mohamed, “Variational inference with normalizing flows,” in International conference on machine learning. PMLR, 2015, pp. 1530–1538.

[84] A. Oord, Y. Li, I. Babuschkin, K. Simonyan, O. Vinyals, K. Kavukcuoglu, G. Driessche, E. Lockhart, L. Cobo, F. Stimberg et al., “Parallel wavenet: Fast high-fidelity speech synthesis,” in International conference on machine learning. PMLR, 2018, pp. 3918–3926.

[85] A. Van den Oord, N. Kalchbrenner, L. Espeholt, O. Vinyals, A. Graves et al., “Conditional image generation with pixelcnn decoders,” Advances in neural information processing systems, vol. 29, 2016.

[86] S. Chen and R. Gopinath, “Gaussianization,” in Advances in Neural Information Processing Systems, T. Leen, T. Dietterich, and V. Tresp, Eds., vol. 13. MIT Press, 2000. [Online]. Available: https://proceedings.neurips.cc/paper/ 2000/file/3c947bc2f7ff007b86a9428b74654de5-Paper.pdf

[87] E. G. Tabak and E. Vanden-Eijnden, “Density estimation by dual ascent of the log-likelihood,” Communications in Mathematical Sciences, vol. 8, no. 1, pp. 217–233, 2010.

[88] E. G. Tabak and C. V. Turner, “A family of nonparametric density estimation algorithms,” Communications on Pure and Applied Mathematics, vol. 66, no. 2, pp. 145–164, 2013.

[89] I. Kobyzev, S. J. Prince, and M. A. Brubaker, “Normalizing flows: An introduction and review of current methods,” IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 11, pp. 3964–3979, 2020.

[90] L. Dinh, D. Krueger, and Y. Bengio, “Nice: Non-linear independent components estimation,” arXiv preprint arXiv:1410.8516, 2014.

[91] L. Dinh, J. Sohl-Dickstein, and S. Bengio, “Density estimation using real nvp,” arXiv preprint arXiv:1605.08803, 2016.

[92] D. P. Kingma, T. Salimans, R. Jozefowicz, X. Chen, I. Sutskever, and M. Welling, “Improved variational inference with inverse autoregressive flow,” Advances in neural information processing systems, vol. 29, 2016.

[93] G. Papamakarios, T. Pavlakou, and I. Murray, “Masked autoregressive flow for density estimation,” Advances in neural information processing systems, vol. 30, 2017.

[94] C.-W. Huang, D. Krueger, A. Lacoste, and A. Courville, “Neural autoregressive flows,” in International Conference on Machine Learning. PMLR, 2018, pp. 2078–2087.

[95] Y. Lu, A. Zhong, Q. Li, and B. Dong, “Beyond finite layer neural networks: Bridging deep architectures and numerical differential equations,” in International Conference on Machine Learning. PMLR, 2018, pp. 3276–3285.

[96] R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural ordinary differential equations,” Advances in neural information processing systems, vol. 31, 2018.

[97] W. Grathwohl, R. T. Chen, J. Bettencourt, I. Sutskever, and D. Duvenaud, “Ffjord: Free-form continuous dynamics for scalable reversible generative models,” arXiv preprint arXiv:1810.01367, 2018.

[98] C. Finlay, J.-H. Jacobsen, L. Nurbekyan, and A. Oberman, “How to train your neural ode: the world of jacobian and kinetic regularization,” in International conference on machine learning. PMLR, 2020, pp. 3154–3164.

[99] N. Rozen, A. Grover, M. Nickel, and Y. Lipman, “Moser flow: Divergence-based generative modeling on manifolds,” Advances in Neural Information Processing Systems, vol. 34, pp. 17 669–17 680, 2021.

[100] Y. Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” Advances in neural information processing systems, vol. 32, 2019.

[101] L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey of methods and applications,” ACM Computing Surveys, vol. 56, no. 4, pp. 1–39, 2023.

[102] Y. Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” 2023.

[103] E. Bengio, M. Jain, M. Korablyov, D. Precup, and Y. Bengio, “Flow network based generative models for non-iterative diverse candidate generation,” Advances in Neural Information Processing Systems, vol. 34, pp. 27 381–27 394, 2021.

[104] Y. Li, S. Luo, H. Wang, and J. Hao, “Cflownets: Continuous control with generative flow networks,” arXiv preprint arXiv:2303.02430, 2023.

[105] L. M. Brunswic, Y. Xu, Y. Feng, M. Lizhuang, and S. Jui, “Nonacyclic gflownets,” arXiv, 2023.

[106] S. Lahlou, T. Deleu, P. Lemos, D. Zhang, A. Volokhova, A. Hernandez-Garc ´ ´ıa, L. N. Ezzine, Y. Bengio, and N. Malkin, “A theory of continuous generative flow networks,” arXiv preprint arXiv:2301.12594, 2023.

[107] N. Malkin, M. Jain, E. Bengio, C. Sun, and Y. Bengio, “Trajectory balance: Improved credit assignment in gflownets,” arXiv preprint arXiv:2201.13259, 2022.

[108] L. Pan, N. Malkin, D. Zhang, and Y. Bengio, “Better training of gflownets with local credit and incomplete trajectories,” arXiv preprint arXiv:2302.01687, 2023.

[109] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017.

[110] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog, vol. 1, no. 8, p. 9, 2019.

[111] R. Child, S. Gray, A. Radford, and I. Sutskever, “Generating long sequences with sparse transformers,” arXiv preprint arXiv:1904.10509, 2019.

[112] Z. Yang, Z. Dai, Y. Yang, J. Carbonell, R. R. Salakhutdinov, and Q. V. Le, “Xlnet: Generalized autoregressive pretraining for language understanding,” Advances in neural information processing systems, vol. 32, 2019.

[113] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pretraining of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018.

[114] A. Agarwal, S. Kakade, and L. F. Yang, “Model-based reinforcement learning with a generative model is minimax optimal,” in Conference on Learning Theory. PMLR, 2020, pp. 67–83.

[115] A. Tirinzoni, R. Poiani, and M. Restelli, “Sequential transfer in reinforcement learning with a generative model,” in International Conference on Machine Learning. PMLR, 2020, pp. 9481–9492.

[116] G. Li, Y. Wei, Y. Chi, Y. Gu, and Y. Chen, “Breaking the sample size barrier in model-based reinforcement learning with a generative model,” Advances in neural information processing systems, vol. 33, pp. 12 861–12 872, 2020.

[117] D. Wang, A. Sundaram, R. Kothari, A. Kapoor, and M. Roetteler, “Quantum algorithms for reinforcement learning with a generative model,” in International Conference on Machine Learning. PMLR, 2021, pp. 10 916–10 926.

[118] M. G. Azar, R. Munos, and B. Kappen, “On the sample complexity of reinforcement learning with a generative model,” arXiv preprint arXiv:1206.6461, 2012.

[119] Y. Du and I. Mordatch, “Implicit generation and modeling with energy based models,” Advances in Neural Information Processing Systems, vol. 32, 2019.

[120] M. Liu, T. He, M. Xu, and W. Zhang, “Energy-based imitation learning,” in Proceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems, 2021, pp. 809–817.

[121] T. Haarnoja, H. Tang, P. Abbeel, and S. Levine, “Reinforcement learning with deep energy-based policies,” in International conference on machine learning. PMLR, 2017, pp. 1352–1361.

[122] L. Kong, J. Cui, Y. Zhuang, R. Feng, B. A. Prakash, and C. Zhang, “End-to-end stochastic optimization with energy-based model,” Advances in Neural Information Processing Systems, vol. 35, pp. 11 341–11 354, 2022.

[123] N. Tagasovska, N. C. Frey, A. Loukas, I. Hotzel, J. Lafrance- ¨ Vanasse, R. L. Kelly, Y. Wu, A. Rajpal, R. Bonneau, K. Cho et al., “A pareto-optimal compositional energy-based model for sampling and optimization of protein sequences,” arXiv preprint arXiv:2210.10838, 2022.

[124] Y. Zhao, J. Xie, and P. Li, “Learning energy-based generative models via coarse-to-fine expanding and sampling,” in International Conference on Learning Representations, 2020.

[125] Y. Li, J. Song, and S. Ermon, “Infogail: Interpretable imitation learning from visual demonstrations,” Advances in neural information processing systems, vol. 30, 2017.

[126] N. Baram, O. Anschel, I. Caspi, and S. Mannor, “End-to-end differentiable adversarial imitation learning,” in International Conference on Machine Learning. PMLR, 2017, pp. 390–399.

[127] P. Geiger and C.-N. Straehle, “Fail-safe adversarial generative imitation learning,” Transactions on Machine Learning Research, 2022.

[128] Y. Wang, C. Xu, B. Du, and H. Lee, “Learning to weight imperfect demonstrations,” in International Conference on Machine Learning. PMLR, 2021, pp. 10 961–10 970.

[129] Y.-H. Wu, N. Charoenphakdee, H. Bao, V. Tangkaratt, and M. Sugiyama, “Imitation learning from imperfect demonstration,” in International Conference on Machine Learning. PMLR, 2019, pp. 6818–6827.

[130] J. Fu, K. Luo, and S. Levine, “Learning robust rewards with adverserial inverse reinforcement learning,” in International Conference on Learning Representations, 2018.

[131] P. Wang, D. Liu, J. Chen, H. Li, and C.-Y. Chan, “Decision making for autonomous driving via augmented adversarial inverse reinforcement learning,” in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 1036–1042.

[132] H. Xiao, M. Herman, J. Wagner, S. Ziesche, J. Etesami, and T. H. Linh, “Wasserstein adversarial imitation learning,” arXiv preprint arXiv:1906.08113, 2019.

[133] J. Sun, L. Yu, P. Dong, B. Lu, and B. Zhou, “Adversarial inverse reinforcement learning with self-attention dynamics model,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 1880–1886, 2021.

[134] V. Huang, T. Ley, M. Vlachou-Konchylaki, and W. Hu, “Enhanced experience replay generation for efficient reinforcement learning,” arXiv preprint arXiv:1705.08245, 2017.

[135] D. Cho, D. Shim, and H. J. Kim, “S2p: State-conditioned image synthesis for data augmentation in offline reinforcement learning,” Advances in Neural Information Processing Systems, vol. 35, pp. 11 534–11 546, 2022.

[136] C. He, S. Huang, R. Cheng, K. C. Tan, and Y. Jin, “Evolutionary multiobjective optimization driven by generative adversarial networks (gans),” IEEE transactions on cybernetics, vol. 51, no. 6, pp. 3129–3142, 2020.

[137] E.-A. Sim, S. Lee, J. Oh, and J. Lee, “Gans and dcgans for generation of topology optimization validation curve through clustering analysis,” Advances in Engineering Software, vol. 152, p. 102957, 2021.

[138] P. R. Kalehbasti, M. D. Lepech, and S. S. Pandher, “Augmenting high-dimensional nonlinear optimization with conditional gans,” in Proceedings of the Genetic and Evolutionary Computation Conference Companion, 2021, pp. 1879–1880.

[139] A. Mandlekar, D. Xu, R. Mart´ın-Mart´ın, S. Savarese, and L. Fei-Fei, “Learning to generalize across long-horizon tasks from human demonstrations,” arXiv preprint arXiv:2003.06085, 2020.

[140] O. Mees, L. Hermann, and W. Burgard, “What matters in language conditioned robotic imitation learning over unstructured data,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11 205–11 212, 2022.

[141] C. Lynch, M. Khansari, T. Xiao, V. Kumar, J. Tompson, S. Levine, and P. Sermanet, “Learning latent plans from play,” in Conference on robot learning. PMLR, 2020, pp. 1113–1132.

[142] E. Rosete-Beas, O. Mees, G. Kalweit, J. Boedecker, and W. Burgard, “Latent plans for task-agnostic offline reinforcement learning,” in Conference on Robot Learning. PMLR, 2023, pp. 1838–1849.

[143] A. Ajay, A. Kumar, P. Agrawal, S. Levine, and O. Nachum, “Opal: Offline primitive discovery for accelerating offline reinforcement learning,” in International Conference on Learning Representations, 2020.

[144] F. Liu, H. Liu, A. Grover, and P. Abbeel, “Masked autoencoding for scalable and generalizable decision making,” Advances in Neural Information Processing Systems, vol. 35, pp. 12 608–12 618, 2022.

[145] J. Han and J. Kim, “Selective data augmentation for improving the performance of offline reinforcement learning,” in 2022 22nd International Conference on Control, Automation and Systems (IC-CAS). IEEE, 2022, pp. 222–226.

[146] W. Xing, J. Lee, C. Liu, and S. Zhu, “Bayesian optimization with hidden constraints via latent decision models,” arXiv preprint arXiv:2310.18449, 2023.

[147] A. Hottung, B. Bhandari, and K. Tierney, “Learning a latent search space for routing problems using variational autoencoders,” in International Conference on Learning Representations, 2021.

[148] P. N. Ward, A. Smofsky, and A. J. Bose, “Improving exploration in soft-actor-critic with normalizing flows policies,” arXiv preprint arXiv:1906.02771, 2019.

[149] D. Akimov, V. Kurenkov, A. Nikulin, D. Tarasov, and S. Kolesnikov, “Let offline rl flow: Training conservative agents in the latent space of normalizing flows,” in 3rd Offline RL Workshop: Offline RL as a”Launchpad”, 2022.

[150] Q. Zheng, M. Le, N. Shaul, Y. Lipman, A. Grover, and R. T. Chen, “Guided flows for generative modeling and decision making,” arXiv preprint arXiv:2311.13443, 2023.

[151] M. Gabrie, G. M. Rotskoff, and E. Vanden-Eijnden, “Adaptive ´ monte carlo augmented with normalizing flows,” Proceedings of the National Academy of Sciences, vol. 119, no. 10, p. e2109420119, 2022.

[152] T. Pearce, T. Rashid, A. Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V. Macua, S. Z. Tan, I. Momennejad, K. Hofmann et al., “Imitating human behaviour with diffusion models,” in The Eleventh International Conference on Learning Representations, 2022.

[153] M. Janner, Y. Du, J. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” in International Conference on Machine Learning. PMLR, 2022, pp. 9902–9915.

[154] A. Ajay, Y. Du, A. Gupta, J. B. Tenenbaum, T. S. Jaakkola, and P. Agrawal, “Is conditional generative modeling all you need for decision making?” in The Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=sP1fo2K9DFG

[155] S. Zhao and A. Grover, “Decision stacks: Flexible reinforcement learning via modular generative models,” Advances in Neural Information Processing Systems, vol. 36, 2024.

[156] Z. Wang, J. J. Hunt, and M. Zhou, “Diffusion policies as an expressive policy class for offline reinforcement learning,” in The Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id= AHvFDPi-FA

[157] C. Chi, S. Feng, Y. Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” arXiv preprint arXiv:2303.04137, 2023.

[158] H. Chen, C. Lu, C. Ying, H. Su, and J. Zhu, “Offline reinforcement learning via high-fidelity generative behavior modeling,” in The Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview. net/forum?id=42zs3qa2kpy

[159] Y. Du, S. Yang, B. Dai, H. Dai, O. Nachum, J. Tenenbaum, D. Schuurmans, and P. Abbeel, “Learning universal policies via text-guided video generation,” Advances in Neural Information Processing Systems, vol. 36, 2024.

[160] Z. Liang, Y. Mu, M. Ding, F. Ni, M. Tomizuka, and P. Luo, “Adaptdiffuser: Diffusion models as adaptive self-evolving planners,” in International Conference on Machine Learning. PMLR, 2023, pp. 20 725–20 745.

[161] L. Yang, Z. Huang, F. Lei, Y. Zhong, Y. Yang, C. Fang, S. Wen, B. Zhou, and Z. Lin, “Policy representation via diffusion probability model for reinforcement learning,” arXiv preprint arXiv:2305.13122, 2023.

[162] H. He, C. Bai, K. Xu, Z. Yang, W. Zhang, D. Wang, B. Zhao, and X. Li, “Diffusion model is an effective planner and data synthesizer for multi-task reinforcement learning,” Advances in neural information processing systems, vol. 36, 2024.

[163] Z. Chen, S. Kiami, A. Gupta, and V. Kumar, “Genaug: Retargeting behaviors to unseen situations via generative augmentation,” arXiv preprint arXiv:2302.06671, 2023.

[164] C. Lu, P. Ball, Y. W. Teh, and J. Parker-Holder, “Synthetic experience replay,” Advances in Neural Information Processing Systems, vol. 36, 2024.

[165] S. Krishnamoorthy, S. M. Mashkaria, and A. Grover, “Diffusion models for black-box optimization,” in International Conference on Machine Learning. PMLR, 2023, pp. 17 842–17 857.

[166] Z. Li, H. Yuan, K. Huang, C. Ni, Y. Ye, M. Chen, and M. Wang, “Diffusion model for data-driven black-box optimization,” arXiv preprint arXiv:2403.13219, 2024.

[167] L. Kong, Y. Du, W. Mu, K. Neklyudov, V. De Bortol, H. Wang, D. Wu, A. Ferber, Y.-A. Ma, C. P. Gomes et al., “Diffusion models as constrained samplers for optimization with unknown constraints,” arXiv preprint arXiv:2402.18012, 2024.

[168] L. Pan, D. Zhang, M. Jain, L. Huang, and Y. Bengio, “Stochastic generative flow networks,” in Uncertainty in Artificial Intelligence. PMLR, 2023, pp. 1628–1638.

[169] L. Pan, D. Zhang, A. Courville, L. Huang, and Y. Bengio, “Generative augmented flow networks,” in The Eleventh International Conference on Learning Representations, 2022.

[170] M. Jiralerspong, B. Sun, D. Vucetic, T. Zhang, Y. Bengio, G. Gidel, and N. Malkin, “Expected flow networks in stochastic environments and two-player zero-sum games,” in The Twelfth International Conference on Learning Representations, 2023.

[171] L. Brunswic, Y. Li, Y. Xu, Y. Feng, S. Jui, and L. Ma, “A theory of non-acyclic generative flow networks,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 10, 2024, pp. 11 124–11 131.

[172] D. Zhang, R. T. Chen, N. Malkin, and Y. Bengio, “Unifying generative models with gflownets,” arXiv preprint arXiv:2209.02606, 2022.

[173] D. Zhang, N. Malkin, Z. Liu, A. Volokhova, A. Courville, and Y. Bengio, “Generative flow networks for discrete probabilistic modeling,” in International Conference on Machine Learning. PMLR, 2022, pp. 26 412–26 428.

[174] M. Kim, S. Choi, H. Kim, J. Son, J. Park, and Y. Bengio, “Ant colony sampling with gflownets for combinatorial optimization,” arXiv preprint arXiv:2403.07041, 2024.

[175] M. Jain, S. C. Raparthy, A. Hernandez-Garcıa, J. Rector-Brooks, ´ Y. Bengio, S. Miret, and E. Bengio, “Multi-objective gflownets,” in International conference on machine learning. PMLR, 2023, pp. 14 631–14 653.

[176] L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch, “Decision transformer: Reinforcement learning via sequence modeling,” Advances in neural information processing systems, vol. 34, pp. 15 084–15 097, 2021.

[177] M. Janner, Q. Li, and S. Levine, “Offline reinforcement learning as one big sequence modeling problem,” Advances in neural information processing systems, vol. 34, pp. 1273–1286, 2021.

[178] Q. Zheng, A. Zhang, and A. Grover, “Online decision transformer,” in international conference on machine learning. PMLR, 2022, pp. 27 042–27 059.

[179] S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gimenez, Y. Sulsky, J. Kay, J. T. Springenberg et al., “A generalist agent,” arXiv preprint arXiv:2205.06175, 2022.

[180] K.-H. Lee, O. Nachum, S. Yang, L. Lee, C. D. Freeman, S. Guadarrama, I. Fischer, W. Xu, E. Jang, H. Michalewski, and I. Mordatch, “Multi-game decision transformers,” in Advances in Neural Information Processing Systems, A. H. Oh, A. Agarwal, D. Belgrave, and K. Cho, Eds., 2022. [Online]. Available: https://openreview.net/forum?id=0gouO5saq6K

[181] B. Zhu, M. Dang, and A. Grover, “Scaling pareto-efficient decision making via offline multi-objective RL,” in The Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=Ki4ocDm364

[182] K. Wang, H. Zhao, X. Luo, K. Ren, W. Zhang, and D. Li, “Bootstrapped transformer for offline reinforcement learning,” Advances in Neural Information Processing Systems, vol. 35, pp. 34 748–34 761, 2022.

[183] H. M. So, L. Bose, P. Dudek, and G. Wetzstein, “Pixelrnn: in-pixel recurrent neural networks for end-to-end-optimized perception with neural sensors,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 25 233–25 244.

[184] S. M. Mashkaria, S. Krishnamoorthy, and A. Grover, “Generative pretraining for black-box optimization,” in International Conference on Machine Learning. PMLR, 2023, pp. 24 173–24 197.

[185] T. Nguyen and A. Grover, “Transformer neural processes: Uncertainty-aware meta learning via sequence modeling,” in International Conference on Machine Learning. PMLR, 2022, pp. 16 569–16 594.

[186] E. Hazan, S. Kakade, and K. Singh, “The nonstochastic control problem,” in Algorithmic Learning Theory, 2020.

[187] X. Chen and E. Hazan, “Black-box control for linear dynamical systems,” in Conference on Learning Theory, 2021.

[188] G. Xiang, S. Li, F. Shuang, F. Gao, and X. Yuan, “Sc-airl: Sharecritic in adversarial inverse reinforcement learning for longhorizon task,” IEEE Robotics and Automation Letters, vol. 9, no. 4, pp. 3179–3186, 2024.

[189] T. Pearce, T. Rashid, A. Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V. Macua, S. Z. Tan, I. Momennejad, K. Hofmann et al., “Imitating human behaviour with diffusion models,” arXiv preprint arXiv:2301.10677, 2023.

[190] C. Kim, J. Kim, Y. Seo, K. Lee, H. Lee, and J. Shin, “Dynamicsaugmented decision transformer for offline dynamics generalization,” in 3rd Offline RL Workshop: Offline RL as a”Launchpad”, 2022.

[191] M. Xu, Y. Shen, S. Zhang, Y. Lu, D. Zhao, J. Tenenbaum, and C. Gan, “Prompting decision transformer for few-shot policy generalization,” in international conference on machine learning. PMLR, 2022, pp. 24 631–24 645.

[192] S. Wang, X. Chen, D. Jannach, and L. Yao, “Causal decision transformer for recommender systems via offline reinforcement learning,” in Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2023, pp. 1599–1608.

[193] S. Emmons, B. Eysenbach, I. Kostrikov, and S. Levine, “Rvs: What is essential for offline rl via supervised learning?” in International Conference on Learning Representations, 2021.

[194] A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu et al., “Rt-1: Robotics transformer for real-world control at scale,” arXiv preprint arXiv:2212.06817, 2022.

[195] S. Sinha, A. Mandlekar, and A. Garg, “S4rl: Surprisingly simple self-supervision for offline reinforcement learning in robotics,” in Conference on Robot Learning. PMLR, 2022, pp. 907–917.

[196] M. Laskin, K. Lee, A. Stooke, L. Pinto, P. Abbeel, and A. Srinivas, “Reinforcement learning with augmented data,” Advances in neural information processing systems, vol. 33, pp. 19 884–19 895, 2020.

[197] T. Ashuach, M. I. Gabitto, R. V. Koodli, G.-A. Saldi, M. I. Jordan, and N. Yosef, “Multivi: deep generative model for the integration of multimodal data,” Nature Methods, pp. 1–10, 2023.

[198] M. Thomas, A. Bender, and C. de Graaf, “Integrating structurebased approaches in generative molecular design,” Current Opinion in Structural Biology, vol. 79, p. 102559, 2023.

[199] F. Kong, J. Li, B. Jiang, H. Wang, and H. Song, “Integrated generative model for industrial anomaly detection via bidirectional lstm and attention mechanism,” IEEE Transactions on Industrial Informatics, vol. 19, no. 1, pp. 541–550, 2021.

[200] K. Minoura, K. Abe, H. Nam, H. Nishikawa, and T. Shimamura, “A mixture-of-experts deep generative model for integrated analysis of single-cell multiomics data,” Cell reports methods, vol. 1, no. 5, 2021.

[201] T. Muller, B. McWilliams, F. Rousselle, M. Gross, and J. Nov ¨ ak, ´ “Neural importance sampling,” ACM Transactions on Graphics (ToG), vol. 38, no. 5, pp. 1–19, 2019.

[202] T. Yu, T. Xiao, A. Stone, J. Tompson, A. Brohan, S. Wang, J. Singh, C. Tan, J. Peralta, B. Ichter et al., “Scaling robot learning with semantically imagined experience,” arXiv preprint arXiv:2302.11550, 2023.

[203] D. Zhang, Y. Zhang, J. Gu, R. Zhang, J. Susskind, N. Jaitly, and S. Zhai, “Improving gflownets for text-to-image diffusion alignment,” arXiv preprint arXiv:2406.00633, 2024.

[204] X. Qiao, Y. Yang, and H. Li, “Defending neural backdoors via generative distribution modeling,” Advances in neural information processing systems, vol. 32, 2019.

[205] T. Lattimore, C. Szepesvari, and G. Weisz, “Learning with good feature representations in bandits and in rl with a generative model,” in International Conference on Machine Learning. PMLR, 2020, pp. 5662–5670.

[206] G. E. Hinton, “Training products of experts by minimizing contrastive divergence,” Neural computation, vol. 14, no. 8, pp. 1771– 1800, 2002.

[207] D. Zhang, H. Dai, N. Malkin, A. C. Courville, Y. Bengio, and L. Pan, “Let the flows tell: Solving graph combinatorial problems with gflownets,” Advances in Neural Information Processing Systems, vol. 36, 2024.

[208] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel, “Benchmarking deep reinforcement learning for continuous control,” in International conference on machine learning. PMLR, 2016, pp. 1329–1338.

[209] S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training of deep visuomotor policies,” The Journal of Machine Learning Research, vol. 17, no. 1, pp. 1334–1373, 2016.

[210] R. Bailo, M. Bongini, J. A. Carrillo, and D. Kalise, “Optimal consensus control of the cucker-smale model,” IFAC-PapersOnLine, vol. 51, no. 13, pp. 1–6, 2018.

[211] A. Cartea, S. Jaimungal, and J. Ricci, “Algorithmic trading, stochastic control, and mutually exciting processes,” SIAM review, vol. 60, no. 3, pp. 673–703, 2018.

[212] H. Wang and X. Y. Zhou, “Continuous-time mean–variance portfolio selection: A reinforcement learning framework,” Mathematical Finance, vol. 30, no. 4, pp. 1273–1308, 2020.

[213] S. Hofer, K. Bekris, A. Handa, J. C. Gamboa, M. Mozifian, ¨ F. Golemo, C. Atkeson, D. Fox, K. Goldberg, J. Leonard et al., “Sim2real in robotics and automation: Applications and challenges,” IEEE transactions on automation science and engineering, 2021.

[214] X. Chen, J. Hu, C. Jin, L. Li, and L. Wang, “Understanding domain randomization for sim-to-real transfer,” in International Conference on Learning Representations, 2022.

[215] J. Hu, H. Zhong, C. Jin, and L. Wang, “Provable sim-to-real transfer in continuous domain with partial observations,” arXiv preprint arXiv:2210.15598, 2022.

[216] W. Xu, S. Keshmiri, and G. Wang, “Adversarially approximated autoencoder for image generation and manipulation,” IEEE Transactions on Multimedia, vol. 21, no. 9, pp. 2387–2396, 2019.

[217] D. Zhang, R. T. Chen, C.-H. Liu, A. Courville, and Y. Bengio, “Diffusion generative flow samplers: Improving learning signals through partial trajectory optimization,” in The Twelfth International Conference on Learning Representations, 2023.

[218] G. B. Margolis, G. Yang, K. Paigwar, T. Chen, and P. Agrawal, “Rapid locomotion via reinforcement learning,” The International Journal of Robotics Research, vol. 43, no. 4, pp. 572–587, 2024.

[219] Y. Li, L. Zhang, and Z. Liu, “Multi-objective de novo drug design with conditional graph generative model,” Journal of cheminformatics, vol. 10, pp. 1–24, 2018.

[220] A. Grover, A. Zweig, and S. Ermon, “Graphite: Iterative generative modeling of graphs,” in International conference on machine learning. PMLR, 2019, pp. 2434–2444.

[221] S. Zhong and J. Ghosh, “Generative model-based document clustering: a comparative study,” Knowledge and Information Systems, vol. 8, pp. 374–384, 2005.

[222] M. Jain, E. Bengio, A. Hernandez-Garcia, J. Rector-Brooks, B. F. Dossou, C. A. Ekbote, J. Fu, T. Zhang, M. Kilgour, D. Zhang et al., “Biological sequence design with gflownets,” in International Conference on Machine Learning. PMLR, 2022, pp. 9786–9801.

[223] M. Kim, T. Yun, E. Bengio, D. Zhang, Y. Bengio, S. Ahn, and J. Park, “Local search gflownets,” in The Twelfth International Conference on Learning Representations, 2023.

[224] X. Xin, T. Pimentel, A. Karatzoglou, P. Ren, K. Christakopoulou, and Z. Ren, “Rethinking reinforcement learning for recommendation: A prompt perspective,” in Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2022, pp. 1347–1357.

[225] M. Deng, J. Wang, C.-P. Hsieh, Y. Wang, H. Guo, T. Shu, M. Song, E. P. Xing, and Z. Hu, “Rlprompt: Optimizing discrete text prompts with reinforcement learning,” arXiv preprint arXiv:2205.12548, 2022.

[226] T. Zhang, X. Wang, D. Zhou, D. Schuurmans, and J. E. Gonzalez, “Tempera: Test-time prompt editing via reinforcement learning,” in The Eleventh International Conference on Learning Representations, 2023.

[227] K. Shao, Z. Tang, Y. Zhu, N. Li, and D. Zhao, “A survey of deep reinforcement learning in video games,” arXiv preprint arXiv:1912.10944, 2019.

[228] L. Chen, X. Hu, W. Tian, H. Wang, D. Cao, and F.-Y. Wang, “Parallel planning: A new motion planning framework for autonomous driving,” IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 1, pp. 236–246, 2018.

[229] W. Si, T. Wei, and C. Liu, “Agen: Adaptable generative prediction networks for autonomous driving,” in 2019 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2019, pp. 281–286.

[230] A. Marathe, D. Ramanan, R. Walambe, and K. Kotecha, “Wedge: A multi-weather autonomous driving dataset built from generative vision-language models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3317–3326.

[231] H. Arnelid, E. L. Zec, and N. Mohammadiha, “Recurrent conditional generative adversarial networks for autonomous driving sensor modelling,” in 2019 IEEE Intelligent transportation systems conference (ITSC). IEEE, 2019, pp. 1613–1618.

[232] J. Huang, S. Xie, J. Sun, Q. Ma, C. Liu, D. Lin, and B. Zhou, “Learning a decision module by imitating driver’s control behaviors,” in Conference on Robot Learning. PMLR, 2021, pp. 1–10.

[233] S. Ding, Y. Tian, F. Xu, Q. Li, and S. Zhong, “Trojan attack on deep generative models in autonomous driving,” in Security and Privacy in Communication Networks: 15th EAI International Conference, SecureComm 2019, Orlando, FL, USA, October 23-25, 2019, Proceedings, Part I 15. Springer, 2019, pp. 299–318.

[234] L. Feng, Q. Li, Z. Peng, S. Tan, and B. Zhou, “Trafficgen: Learning to generate diverse and realistic traffic scenarios,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 3567–3575.

[235] F. Lateef, M. Kas, and Y. Ruichek, “Saliency heat-map as visual attention for autonomous driving using generative adversarial network (gan),” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 5360–5373, 2021.

[236] J. Sun, Y. Jiang, J. Qiu, P. Nobel, M. J. Kochenderfer, and M. Schwager, “Conformal prediction for uncertainty-aware planning with diffusion dynamics model,” Advances in Neural Information Processing Systems, vol. 36, 2024.

[237] A. Ghosh, B. Bhattacharya, and S. B. R. Chowdhury, “Sadgan: Synthetic autonomous driving using generative adversarial networks,” arXiv preprint arXiv:1611.08788, 2016.

[238] A. Grover, “Generative decision making under uncertainty,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 13, 2023, pp. 15 440–15 440.

[239] D. Zhang, H. Dai, N. Malkin, A. Courville, Y. Bengio, and L. Pan, “Let the flows tell: Solving graph combinatorial optimization problems with gflownets,” arXiv preprint arXiv:2305.17010, 2023.

[240] Z. Sun and Y. Yang, “Difusco: Graph-based diffusion solvers for combinatorial optimization,” Advances in Neural Information Processing Systems, vol. 36, 2024.

[241] Y. Bengio, A. Lodi, and A. Prouvost, “Machine learning for combinatorial optimization: a methodological tour d’horizon,” European Journal of Operational Research, vol. 290, no. 2, pp. 405– 421, 2021.

[242] J. Zhang, C. Liu, X. Li, H.-L. Zhen, M. Yuan, Y. Li, and J. Yan, “A survey for solving mixed integer programming via machine learning,” Neurocomputing, vol. 519, pp. 205–217, 2023.

[243] E. Khalil, H. Dai, Y. Zhang, B. Dilkina, and L. Song, “Learning combinatorial optimization algorithms over graphs,” Advances in neural information processing systems, vol. 30, 2017.

[244] W. Kool, H. Van Hoof, and M. Welling, “Attention, learn to solve routing problems!” arXiv preprint arXiv:1803.08475, 2018.

[245] Y.-D. Kwon, J. Choo, B. Kim, I. Yoon, Y. Gwon, and S. Min, “Pomo: Policy optimization with multiple optima for reinforcement learning,” Advances in Neural Information Processing Systems, vol. 33, pp. 21 188–21 198, 2020.

[246] M. Kim, J. Park, and J. Park, “Sym-nco: Leveraging symmetricity for neural combinatorial optimization,” Advances in Neural Information Processing Systems, vol. 35, pp. 1936–1949, 2022.

[247] C. K. Joshi, T. Laurent, and X. Bresson, “An efficient graph convolutional network technique for the travelling salesman problem,” arXiv preprint arXiv:1906.01227, 2019.

[248] Z.-H. Fu, K.-B. Qiu, and H. Zha, “Generalize a small pre-trained model to arbitrarily large tsp instances,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 8, 2021, pp. 7474–7482.

[249] R. Qiu, Z. Sun, and Y. Yang, “Dimes: A differentiable meta solver for combinatorial optimization problems,” Advances in Neural Information Processing Systems, vol. 35, pp. 25 531–25 546, 2022.

[250] C. Wang, Z. Yu, S. McAleer, T. Yu, and Y. Yang, “Asp: Learn a universal neural solver!” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.

[251] Z. Sun and Y. Yang, “Difusco: Graph-based diffusion solvers for combinatorial optimization,” Advances in Neural Information Processing Systems, vol. 36, pp. 3706–3731, 2023.

[252] S. Sanokowski, S. Hochreiter, and S. Lehner, “A diffusion model framework for unsupervised neural combinatorial optimization,” arXiv preprint arXiv:2406.01661, 2024.

[253] B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter-efficient prompt tuning,” arXiv preprint arXiv:2104.08691, 2021.

[254] N. Ding, S. Hu, W. Zhao, Y. Chen, Z. Liu, H.-T. Zheng, and M. Sun, “Openprompt: An open-source framework for promptlearning,” arXiv preprint arXiv:2111.01998, 2021.

[255] X. Liu, K. Ji, Y. Fu, W. Tam, Z. Du, Z. Yang, and J. Tang, “Ptuning: Prompt tuning can be comparable to fine-tuning across scales and tasks,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2022, pp. 61–68.

[256] T. Vu, B. Lester, N. Constant, R. Al-Rfou, and D. Cer, “Spot: Better frozen model adaptation through soft prompt transfer,” arXiv preprint arXiv:2110.07904, 2021.

[257] D. Liu, M. Jain, B. F. Dossou, Q. Shen, S. Lahlou, A. Goyal, N. Malkin, C. C. Emezue, D. Zhang, N. Hassen et al., “Gflowout: Dropout with generative flow networks,” in International Conference on Machine Learning. PMLR, 2023, pp. 21 715–21 729.

[258] D. W. Zhang, C. Rainone, M. Peschl, and R. Bondesan, “Robust scheduling with gflownets,” arXiv preprint arXiv:2302.05446, 2023.

[259] A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sutton, S. Gehrmann et al., “Palm: Scaling language modeling with pathways,” Journal of Machine Learning Research, vol. 24, no. 240, pp. 1–113, 2023.

决策中的生成模型：综述 听光Vic

决策中的生成模型：综述听光Vic