Large-Scale Generative Simulation AI (LS-GenAI)

• LS-GenAI is a simulation-centered generative AI paradigm that integrates high-fidelity simulation environments with sequential decision-making for experimental design and rapid adaptation.
• It employs structured state management and query-based context selection to maintain consistency and coherence over long simulation horizons.
• LS-GenAI systems combine multiple components—digital twins, modular generative pipelines, and closed-loop evaluation—to optimize scalability, safety, and task customization.

Large-Scale Generative Simulation AI (LS-GenAI) denotes a simulation-centered extension of generative AI in which models do not merely produce text, images, or other artifacts, but generate, emulate, and interact with high-fidelity simulation environments and sequential decision processes. In the formulation that explicitly introduced the term, LS-GenAI is a “large-scale generative AI paradigm for simulation and sequential decision-making,” framed as a “doubly generative paradigm for simulation and sequential decision-making” that supports meaningful experimental design, fast adaptation of learned skills, counterfactual evaluation, scientific discovery, and task customization for decision-making (Wang et al., 2023). Across later work, the term is used more broadly to describe systems in which generative models become components of simulators, digital twins, network infrastructures, model-generation pipelines, and closed-loop discovery workflows rather than remaining confined to content generation alone (Sunehag et al., 30 Jan 2025).

1. Definition and conceptual boundaries

LS-GenAI is positioned at the intersection of simulation science and artificial intelligence, with the explicit aim of operating in settings where direct experimentation is costly, slow, or impractical. Its motivating problems are “limited learning resources” and “overdependence on scientific discovery empiricism,” and its proposed function is to reduce realistic sampling complexity while accelerating scientific progress, catalyzing discovery, and supporting domains such as interventional clinical research, robotics, and life systems (Wang et al., 2023).

Within this framing, generative AI is not restricted to AI-generated content. The foundational argument is that deep generative models enable data compression, representation disentanglement, and causal inference, including counterfactual prediction of the form $$p\left(y\middle| x,\mathrm{do}\left(z=z_0\right)\right)$$, and that these capabilities can be generalized from language and vision to simulated worlds and decision processes (Wang et al., 2023). A recurrent implication across the literature is that LS-GenAI should be understood as a systems paradigm rather than a single model family: diffusion models, GANs, VAEs, PINNs, transformers, LLMs, multimodal models, and domain-specific hybrids all appear as components of larger simulation workflows rather than as self-sufficient endpoints (Hadid et al., 2024).

A central misconception addressed implicitly by several papers is that LS-GenAI is equivalent to free-form generative chat or synthetic media generation. The literature instead treats it as simulation-driven intelligence: in autonomous driving, GenAI is described as a route toward “simulation-first autonomy” via synthetic data, digital twins, and multimodal world generation; in geoscience, generative models are framed as learned simulators, synthetic data engines, and reasoning assistants; in urban planning, planning itself is reformulated as conditional generation of land-use configurations under geospatial, social, and policy-driven objectives (Wang et al., 13 May 2025). This suggests that LS-GenAI is best defined by its operational role in model-based experimentation, counterfactual analysis, and sequential control, not by the modality of its outputs.

2. Recurring architectural principles

The most explicit programming formalization in the corpus is “Simulation Streams,” which presents a simulation as an explicit state machine whose evolution is driven by a sequence of operators. A Simulation Stream is defined as a tuple $S=(X,O,L,T)$, where $X$ is the state space, $O$ the set of operators, $L$ the LLM sampling function, and $T$ the termination conditions; state evolution is written as $x_{t+1}=F(x_t,o_t,l_t)$, with the next operator chosen by $o_{t+1}=n(o_t)$ (Sunehag et al., 30 Jan 2025). The design goal is “minimal interference”: to preserve LLM strengths in reasoning, planning, dialogue, and open-ended generation while constraining them enough to preserve consistency, enforce world rules, and keep long simulations on track.

Two design ideas in that framework recur widely across LS-GenAI. The first is explicit state management with structured updates. Operators modify one part of the state, optionally invoke the model, and emit a structured row into an output stream, thereby preserving regularity over long horizons. The second is query-based context selection. Instead of exposing the model to a growing unstructured transcript, the framework uses substreams and a query function

$q(c,s,O)=\{o\in O : c(s,o)\text{ is true}\}$

to retrieve only relevant histories, with the stated purpose of keeping context “in-distribution” and supporting coherence over hundreds or thousands of steps (Sunehag et al., 30 Jan 2025).

The broader LS-GenAI literature adds several complementary architectural principles. One is task distribution design: tasks should be representative of broad scenarios, especially for zero-shot and few-shot learning, but exhaustive exploration becomes computationally expensive as problem complexity increases and masked scenarios can grow exponentially with data complexity $\left|\mathcal{X}\right|$ in a combinatorial sense (Wang et al., 2023). Another is geometric inductive bias, especially in scientific and spatial domains, where symmetry and scale separation are treated as essential constraints for generalization and physical plausibility (Wang et al., 2023). A third is multi-view evaluation, with fidelity, diversity, and safety treated as coequal criteria rather than interchangeable proxies (Wang et al., 2023).

Many later systems instantiate the same pattern in domain-specific ways. WorldGen organizes scalable synthetic world production through a three-stage pipeline of loading assets, structural modification and animation, and rendering, with controllable structure, motion, optics, and annotations (Singh et al., 2022). MBSE-oriented work uses “scalable templates” to convert long-form simulation-model creation into constrained code-completion tasks in a formal language, thereby reducing context loss and truncation in model generation (Zhang et al., 9 Mar 2025). MOFA organizes generation, screening, validation, adsorption estimation, and retraining as concurrent stages in an online learning loop rather than a strictly serial pipeline (Yan et al., 18 Jan 2025). Across these examples, LS-GenAI tends to combine generative modules with explicit state, domain constraints, reusable components, and staged validation rather than relying on unconstrained end-to-end sampling.

3. Generative simulators, worlds, and agentic environments

A major branch of LS-GenAI concerns the direct generation of worlds, scenes, and agents. WorldGen is an open-source framework for autonomously generating “countless structured and unstructured 3D photorealistic scenes” such as city view, object collection, and object fragmentation, with rich ground-truth annotations. It is built around Blender Cycles, Bullet physics, OptiX denoising, and Global Illumination, and exposes control over texture, object structure, motion, camera and lens properties, lighting, and weather, while producing outputs such as depth maps, optical flow, stereo images, semantic maps, surface normals, and event frames (Singh et al., 2022). Its optical-flow evaluation on Sintel reports Endpoint Error values of 1.86 on Sintel Clean and 2.87 on Sintel Final for WorldGen-trained RAFT, compared with 2.27/3.76 for FlyingChairs and 2.08/2.75 for AutoFlow, illustrating the claim that photorealistic generative simulation can improve downstream learning (Singh et al., 2022).

LLM-centered agentic simulation appears in Simulation Streams. The paper demonstrates an ongoing market economy simulation with consumers, producers, a market maker, and common resources; a social simulation in which Alice, Bob, and Charlie play catch in a park while a world entity injects events such as ice cream trucks and squirrels; and a suite of classical RL benchmark tasks. The market simulation tracks eight markets—housing, houses, bread, wood, grain, milk, fish, and labor—and reports capital distribution, GDP, utility, supply, demand, and inventory dynamics over 250 steps, with utility defined as $S=(X,O,L,T)$0 when labor inventory exceeds 5, and 0 otherwise (Sunehag et al., 30 Jan 2025). In the social simulation, Gemini-2.0-Flash-Exp achieves perfect consistency across all 250 decisions in the 10-run evaluation, while Gemini-2.0-Flash-Thinking-Exp makes only a few mistakes, demonstrating that exact format discipline and state consistency can be maintained across multiple agents and repeated interactions (Sunehag et al., 30 Jan 2025).

Related robotic systems shift the locus of generation from the world state alone to the full learning problem. Gen2Sim automates generation of 3D assets, task descriptions, temporal decompositions, and Python reward functions using large pre-trained generative models of language and vision, then trains policies with PPO in IsaacGym; its central empirical claim is that it succeeds in learning diverse long-horizon tasks “where reinforcement learning with non temporally decomposed reward functions fails” (Katara et al., 2023). RoboGen similarly advocates a propose-generate-learn cycle in which foundation models propose tasks, generate scenes and supervision, decompose high-level tasks, choose among reinforcement learning, motion planning, and trajectory optimization, and repeatedly produce an “endless stream of skill demonstrations” (Wang et al., 2023). This suggests a broader LS-GenAI pattern in robotics: generative models are used less as policy executors than as generators of curricula, environments, decompositions, and supervision.

4. Scientific, engineering, and domain-specific instantiations

LS-GenAI has been extended into scientific discovery and engineering model construction, where the objective is often not world generation per se but the generation of simulation-ready artifacts coupled to downstream physics or formal models. MOFA is exemplary in this regard. It combines a generative model, MOFLinker, derived from DiffLinker, with a tiered simulation workflow that processes linkers, assembles MOFs, validates structure, optimizes cells with DFT, estimates adsorption using Monte Carlo simulation, and retrains on the best-performing discoveries. The paper reports a 450-node run on Polaris using 14,400 AMD Zen 3 CPU cores and 1,800 NVIDIA A100 GPUs, with 114 MOFs/hour sustained throughput, linear scaling from 32 to 450 nodes, and a top candidate achieving 4.05 mol/kg at 0.1 bar, placing it in the top 5 of the hMOF dataset (Yan et al., 18 Jan 2025). The workflow illustrates LS-GenAI as closed-loop online learning under heterogeneous simulation cost.

In geoscience, generative AI is framed as a foundation layer for synthetic data generation, surrogate simulation, probabilistic forecasting, and domain reasoning. GANs are used for seismic waveform synthesis, geological facies, porous media, remote sensing scenes, and urban morphologies; PINNs encode ODEs and PDEs into the loss to improve physical plausibility; and domain-adapted LLMs such as K2 and vision-LLMs such as GeoChat support geoscience question answering, reasoning, and multimodal interaction (Hadid et al., 2024). The geoscience survey therefore treats LS-GenAI as a unification of learned simulators, physics-aware models, and scientific foundation models, while emphasizing the need for physical grounding and trust.

In model-based systems engineering, generative AI is used to turn product design documents into executable simulation models of physical properties. The proposed workflow employs BERT fine-tuned for NER and sentence classification to construct a simulation model corpus, uses X language as the integrated modeling and simulation representation, and introduces scalable templates for couple, discrete, and continuous classes so that generation becomes modular code completion rather than unconstrained long-form synthesis (Zhang et al., 9 Mar 2025). The reported final scores for template-guided generation—0.810 for CodeGemma, 0.825 for CodeQwen, and 0.870 for DeepSeek-Coder, compared with approximately 0.16 for direct generation—support the claim that engineering structure and domain scaffolding substantially improve simulation-model generation quality (Zhang et al., 9 Mar 2025).

Other domain extensions generalize the same logic. Text-to-image generation is presented as a new capability for modeling and simulation, supporting model explanation, visualization of simulation outcomes, educational materials, and interfaces between heterogeneous models in multi-scale simulations through reproducible local pipelines (Giabbanelli, 18 Jun 2026). Autonomous driving surveys frame GenAI as a simulation engine for images, LiDAR, trajectories, occupancy, video, digital twins, and closed-loop evaluation (Wang et al., 13 May 2025). Urban planning work reformulates planning as conditional generation of land-use configurations and argues for theory-guided generation, digital twins, and human-machine co-design (Fu, 19 Jul 2025). Taken together, these cases indicate that LS-GenAI is not tied to any single scientific substrate; it is a transferable pattern of generative proposal, simulation-based evaluation, and iterative refinement.

5. Systems infrastructure and scaling regimes

A defining characteristic of LS-GenAI is that scalability is a systems problem as much as a modeling problem. The edge-cloud overview frames “GenAI at scale” through three interacting pressures—huge model sizes, massive user request volume, and explosive growth in AI-generated content—and argues that cloud-only deployment is insufficient because of latency, power, communication, privacy, and infrastructure cost (Wang et al., 2023). It presents edge-cloud collaboration as the enabling architecture: cloud for training and heavy generation, edge servers for low-latency inference and personalization, and user devices for lightweight local intelligence. The explicit latency decomposition

$S=(X,O,L,T)$1

is used to explain why cloud-only, edge-only, and edge-cloud systems trade off differently (Wang et al., 2023).

Several LS-GenAI systems make this infrastructure requirement explicit. MOFA coordinates GPU-intensive generation, CPU- and GPU-optimized screening, and retraining through Parsl, Colmena, and ProxyStore, with dynamic reallocation across stages and compute workers spending over 99% of their time doing useful work in the main task types (Yan et al., 18 Jan 2025). Simulation Streams provides a Flask-based web application, command-line execution, and an ECSEditor backend to manage entities, variables, systems definitions, execution, and querying (Sunehag et al., 30 Jan 2025). These details matter because LS-GenAI workflows often involve long horizons, multiple simulators, or mixed CPU/GPU stages that cannot be served by single-model inference abstractions alone.

Networking-oriented work extends the paradigm further by treating generative models as communication infrastructure. In prediction-based networking, a source may transmit a compact prompt to a GenAI-enabled node near the destination, which then generates an approximation of the original content. The paper presents this as a step toward LS-GenAI in which generative models become core infrastructure for machine-to-machine communication, control, sensing, and simulation at network scale, with benefits including throughput expansion and latency reduction; prior work is cited as improving delivered image flow by more than 100% relative to a JPEG baseline (Thorsager et al., 7 Oct 2025). The same paper identifies three enabling directions: initialization protocols and efficient prompt-size selection, implicit prompting for timely delivery of data, and GenAI as an alternative or complement to TCP congestion control (Thorsager et al., 7 Oct 2025).

A related vehicular mixed-reality Metaverse architecture combines digital twins, RSUs, virtual simulators, text-to-image diffusion, and auction-based resource allocation. It is designed to synthesize “unlimited conditioned traffic and driving data” for offline training and online prediction, and reports that generative AI-empowered simulations increase surplus by at least 150% in its market formulation (Xu et al., 2023). A plausible implication is that LS-GenAI is increasingly inseparable from scheduling, offloading, and incentive design once it is deployed as infrastructure rather than as isolated model inference.

6. Evaluation, limitations, and research trajectory

The literature is unusually consistent in treating evaluation as multidimensional. The original LS-GenAI formulation proposes multi-view evaluation systems based on fidelity, diversity, and safety, explicitly rejecting the idea of a universally applicable criterion for generative performance (Wang et al., 2023). Domain surveys repeat this point in more applied terms: autonomous driving work proposes score-card evaluation across distributional fidelity, prompt or layout alignment, temporal consistency, physical realism, downstream transfer, and human preference and safety review; the geoscience survey stresses uncertainty quantification, physical interpretability, and trustworthiness; and MBSE work introduces domain-specific metrics centered on degree of error and model consistency rather than generic code scores (Wang et al., 13 May 2025).

Limitations also recur across domains. Simulation Streams acknowledges an implicit dependence on the model’s ability to follow prescribed formats and remain consistent, sometimes requiring resampling or correction when a model fails (Sunehag et al., 30 Jan 2025). Text-to-image simulation workflows note that repeated image generation can break identity and spatial continuity across calls, making consistency-preserving techniques necessary for multi-image or multi-step simulation use (Giabbanelli, 18 Jun 2026). The autonomous driving survey emphasizes rare-case generalization, budget-limited implementation, evaluation and safety checks, regulatory compliance, ethical concerns, and environmental effects (Wang et al., 13 May 2025). The geoscience survey warns about black-box behavior, biased data, hallucinated outputs, malicious use such as fabricated satellite imagery, computational cost, and carbon footprint (Hadid et al., 2024).

Industrial evidence from large-scale cyber-physical systems sharpens these concerns. A cross-company workshop with 33 participants from six companies identifies three priority research directions: AI-generated scenarios and environment models, simulators and AI in CI/CD pipelines, and trustworthiness in generative AI for simulation (Sadrnezhaad et al., 5 Dec 2025). The reported practitioner concerns include too much data, difficulty generating good scenarios and realistic environments, the need to choose which scenarios are worth running, interoperability problems among simulators, and worries about trust, regulation, traceability, and over-reliance (Sadrnezhaad et al., 5 Dec 2025). This suggests that the principal obstacles to LS-GenAI adoption are no longer limited to raw generative capability; they are increasingly concentrated in workflow integration, validation, governance, and systems interoperability.

The research trajectory implied by the corpus is therefore twofold. First, LS-GenAI is moving from content generation toward explicit world models, digital twins, controllable simulators, and closed-loop optimization. Second, the field is moving from stand-alone model demos toward modular infrastructures that combine generation, simulation, evaluation, and retraining under resource constraints. The resulting agenda includes better world models, stronger geometric and physical priors, task-distribution design, multi-scale and partial-observability handling, trustworthy evaluation, efficient deployment, and tighter integration of generative modules with simulation science and operational systems (Wang et al., 2023).