MarketGen: Generative Market Simulation

• MarketGen is a family of generative simulation systems that model complex market environments using AI-driven advertising, procedural retail scenarios, and financial multi-agent techniques.
• It enables high-fidelity scenario generation with applications in embodied AI for retail, controlled financial market simulation, and strategic game-theoretical ecosystem analysis.
• Researchers leverage MarketGen benchmarks and paradigms to assess metrics like click-through rates, agent success rates, and the fidelity of simulated market equilibria.

MarketGen refers to a family of simulation systems and algorithmic paradigms employing generative models to model, analyze, and interact with complex market environments. This encompasses generative engine marketing with AI-injected advertising, agent-based embodied market simulation, financial market environment simulators, LLM-driven consumer and marketing agent simulations, and generative modeling of market equilibria and games. MarketGen systems typically enable high-fidelity and controllable scenario generation for research in economics, marketing, autonomous retail robotics, and finance.

1. Generative Engine Marketing and Ad-Injected Response Benchmarking

MarketGen in the context of generative engine marketing defines a new advertising paradigm in which LLM-based services (e.g., chatbots, search overviews) monetize user interactions by natively embedding paid promotions within generated text. Unlike classical search engine marketing with “sponsored slots,” the response itself becomes the ad vehicle, without discrete ad panels. The central computational problem is Ad-Injected Response (AIR) generation: producing a single answer that meets user information need and seamlessly embeds ad content sourced from a database $D$ without compromising reading flow or trust (Hu et al., 17 Sep 2025).

The GEM-Bench benchmark operationalizes MarketGen/GEM by providing three datasets (MT-Human, LM-Market, CA-Prod) spanning chatbot and search scenarios, offering thousands of human- and LLM-annotated queries and ads with annotated taxonomies. The benchmark introduces a metric ontology spanning quantitative response/ad flow and coherence and LLM-judged qualitative axes (accuracy, naturalness, trust, noticeability, click propensity).

Its modular multi-agent baseline implements generation, candidate ad retrieval (embedding-based), minimal-disturbance ad placement, and post-hoc context rewriting. Experiments show that prompt-based ad injection yields higher click-through rates but degrades satisfaction and trust, while response-based retrieval plus rewriting improves accuracy, personality, and trust at a moderate engagement penalty and greater compute cost due to supplementary model calls.

Key open problems include: slot-identification for ad suitability, multi-ad response planning, click estimation in unstructured text, and closing the gap between LLM judgment and real-user feedback. GEM-Bench lays foundational datasets, metrics, and architecture for systematic research into MarketGen within AI search/assistant monetization (Hu et al., 17 Sep 2025).

2. MarketGen for Embodied AI: Procedural Supermarket Simulations

Embodied MarketGen systems target commercial environments such as supermarkets by constructing detailed, physically plausible 3D environments using agent-based procedural content generation (PCG) (Hu et al., 26 Nov 2025). MarketGen’s architecture integrates multimodal (text/image) scene specification, a spatial-semantic agent pipeline (SpatialAgent, SemanticAgent, ReflectionRefiner), and a large 3D asset library exceeding 1100 parameterized products and facilities.

The PCG pipeline automatically composes realistic supermarket layouts: zone-partitioned floorplans, semantically appropriate shelf/facility assignment, asset retrieval and placement subject to practical constraints (e.g., shelf clearance, adjacency principles from store design). The reflection refiner uses LLM-driven feedback to optimize scene consistency.

MarketGen enables challenging robotic simulation benchmarks—long-horizon Checkout Unloading for cashier arms and In-Aisle Item Collection for mobile salesperson agents—formulated as finite-horizon MDPs and evaluated by success rate, SPL, and path statistics. Sim-to-real transfer experiments show alignment in grasp success rates between simulated and physical environments.

The platform supports rapid scene diversification (~10 times faster than previous benchmarks), exposes a modular manipulation system (affordance-labeled asset library, SAM, motion planning), and is evaluated with MLLM baselines (GPT-4o, Claude-S-4.5, Gemini-2.5-Pro, Qwen3-VL-Plus), all of which struggle with long-horizon, cluttered tasks. MarketGen provides a scalable standard for embodied AI in retail environments (Hu et al., 26 Nov 2025).

3. MarketGen in Financial Market Simulation and Controllable Generation

MarketGen architectures for financial markets leverage deep generative models to mimic high-fidelity market processes, either directly at the order level or via statistical summaries. These include:

• CGAN-based order generation in ABIDES simulations: Here, a conditional Wasserstein GAN with gradient penalty (WGAN-GP) acts as a world agent, generating order submissions based on recent order-book history, with preservation of stylized facts (return autocorrelation decay, volatility clustering, etc.) and endogenous response to agent interventions (Coletta et al., 2021). Model inputs are aggregate LOB histories, LSTM-encoded, and convolutionally processed to sample orders, ensuring market response realism.
• Signature-based VAE for small data: For environments with limited pathwise data, MarketGen leverages truncated path signatures (universal, parsimonious feature maps) as input to VAEs, permitting stable inference and synthesis even with small sample sizes. Simulation output can be validated by signature-MMD, which metrizes pathwise law similarity (Bühler et al., 2020). The approach is robust against nonstationarity and links generated data directly to down-stream deep hedging performance.
• Contextual MarketGANs (MarketGen): These employ hybrid GAN—supervisor—autoencoder models for time-series generation conditioned on context (market regime, ticker, history). Supervised initialization precedes adversarial training, and evaluations focus not only on fidelity but also semantic attribute alignment, market metric preservation, and downstream utility (forecasting task improvement) (Xia et al., 2023).
• Order-level MarketGen foundation models (MarS): Transformer-based dual-module ensembles learn next-token and batchwise distributions over full historical order streams. The system supports “signal” (scenario/context) conditioning, real-time agent interaction, and market impact learning, with clear evidence of scaling laws and simulation fidelity across multiple stylized fact axes (Li et al., 2024).
• Diffusion-guided “Meta Agent” market generation: A conditional DDPM models the distribution of market-state vectors (returns, order intensities) guided by scenario controls; an economics-prior meta-agent generates plausible microstructure orders given these dynamics. The approach achieves controllability to target indicators and high KL-fidelity on order-flow statistics, and supports downstream RL policy training (Huang et al., 2024).

4. MarketGen in Multi-Agent Marketing and Consumer Behavior Simulation

MarketGen, instantiated as LLM-based sandbox simulations, models granular consumer decision making and emergent marketing phenomena via generative agents (Chu et al., 20 Oct 2025). The system defines:

• Discrete-time world: Agents with internal “chain-of-thought” histories, beliefs, persistent habits, and social network ties, moving across locations, making choices based on LLM-evaluated utility and social influence (peer-pressure term).
• Persona and habit modeling: Agent purchasing behavior is dynamically influenced by memory, observed promotions, and prior habit strength, with endogenous habit formation and decay. Each agent samples possible actions, ranks them, and selects those maximizing short-term utility subject to social context.
• Marketing scenario experimentation: Simulation of price-discount campaigns quantifies conversion rates, revenue effects, visit frequency distributions, and habit/word-of-mouth propagation under varied configurations (price elasticity surfaces, ROI curves).
• Strategy iteration: MarketGen supports systematic A/B testing of campaign designs (e.g., varying discount depth, timing, or channel), with data logging enabling attribution and tuning. The platform highlights memory effects, peer influence, and habit persistence not captured by rule-based agent models or post-event analyses (Chu et al., 20 Oct 2025).

5. MarketGen for Automated Research, Content Synthesis, and KOL Generation

• LLM-Driven Business Analysis and Market Reporting (MaRGen): MarketGen deployments automate the market research life cycle via specialized LLM agents: Researcher (data querying, hypothesis formation), Retriever (few-shot mining of consultant knowledge), Writer (insight composition with embedded code/figures), and Reviewer (LLM-based critique and iterative refinement). The iterative improvement cycle, combined with prompt-tuned in-context learning and structured evaluation (clarlty, layout, pairwise quality), achieves rapid, inexpensive, and analytically robust report generation (Koshkin et al., 2 Aug 2025).
• Modular Generative AI for Branded Content (GenKOL MarketGen): MarketGen as a marketing content engine connects garment generation, makeup transfer, hairstyle editing, and background synthesis through a modular orchestration layer. Each stage is implemented with state-of-the-art U-Net diffusion or GANs, exposing plug-and-play workflows amenable to REST/gRPC scaling across cloud clusters. Evaluation covers FID, user realism assessments, and pipeline throughput metrics, supporting best practices in brand-safe prompt design and iterative previewing (To et al., 18 Sep 2025).
• GenAI Strategy Co-Creation and Explainability (MindFuse): MindFuse-based MarketGen fuses CTR-predictive modeling, persona mining, and attention-based explainability into a co-creative framework that distills content pillars, mines psychographic personas, generates and iterates creatives, and delivers real-time, LLM-guided optimization of active campaigns. Evaluation demonstrates multi-fold efficiency gains and the pivotal role of explainable, metric-driven creative adaptation (Farseev et al., 1 Dec 2025).

6. MarketGen in Game-Theoretical Modeling of Model-Platform-User Ecosystems

MarketGen formalizes generative ecosystem competition as a three-layer market game: model providers, platforms, and heterogeneous users (Wei et al., 19 Feb 2026). Each platform selects a generative model for deployment; user types choose platforms maximizing expected reward from sampled content. The analysis precisely characterizes Nash equilibria in terms of global model quality, localized deviation advantages (attraction to subsegments), and ties conditions for differentiation or homogeneity to observed user preference distributions.

Key results include:

• Equilibrium characterization: Both differentiated and homogeneous equilibria exist depending on the balance between global mean scores and localized segment wins (precisely stated in satisfaction gap inequalities).
• Welfare and diversity: Social planner coverage, user welfare, and model diversity may not always increase with additional models/platforms due to market convergences or cycles.
• Best-response model entry: Providers can optimize for adoption via data reweighting (emphasizing segments with high adoption probability) or direct-gradient adjustments against a Bradley–Terry softmax surface over type-margins.

The framework delivers actionable sufficient conditions for sustainable diversity and “beneficial” competition but also identifies paradoxes (entry reducing welfare, cycles undermining diversity) (Wei et al., 19 Feb 2026).

7. Theoretical and Algorithmic Innovations in MarketGen

Recent MarketGen work pioneers hybrid deep RL–GAN architectures for continuous-time financial market equilibrium computation, termed reinforcement-linked generative adversarial learning (Kratsios et al., 5 Apr 2025). This approach alternates between:

• Generator: Each agent solves a control RL subproblem to maximize utility given current price dynamics, parameterized by a neural SDE.
• Discriminator: Adjusts candidate price drift/volatility to enforce market clearing and terminal constraints, directly feeding its estimate of $\mu_t, \sigma_t$ back to the generator.
• Reinforcement link: Stabilizes the adversarial loop with direct, fine-grained feedback.
• Approximation guarantee: There is a small-time, linear-parameter universal approximation theorem for neural SDE-based FBSDE solvers under mild regularity.

Experiments recover both closed-form and intractable equilibrium asset dynamics, even in the presence of nonlinear frictions and multi-agent interaction, without mode collapse or instability (Kratsios et al., 5 Apr 2025).

---

In summary, MarketGen encompasses a rapidly evolving class of generative modeling systems for market simulation, algorithmic content integration, agent-based retail/marketing experimentation, financial multi-agent simulation, and game-theoretical analysis of generative platform competition. The technical diversity spans transformer LMMs, diffusion models, adversarial nets, multi-agent LLM sandboxes, modular diffusion service orchestration, and multi-level economic analyses. Each class is distinguished by its focus on controllability, fidelity, scenario diversity, downstream utility, and rigorous evaluation—providing platforms and benchmarks foundational to empirical, algorithmic, and theoretical research in modern computational markets.