Tabletop Exercises Are Old School. But Are LLMs Better at War Planning?

How LLMs complement or compete with other quantitative based military planning approaches: Dialogue, Interpretation, Orchestration, and Creation

The Founding Fathers: Simulation, Optimization, Kriegsspiel….and Blinky.

This is the second post on LLMs and military planning. Part 1 can be read here.

Tabletop Exercises Are Old School. But Are LLMs Better at War Planning?

At some point, you come full circle and become the skeptical "old guard." I remember my younger days as a researcher, developing quantitative models for military strategy and tactical planning with the US Air Force at The RAND Corporation. It was an exciting time – we had improved software tools and computing resources, and we could successfully blend classical and modern mathematical techniques. We were convinced that a strong mix of optimization and simulation could yield “better solutions” to any problem. 

Some of the legends in Operations Research that we worked with – with decades of military planning experience – weren’t bought in. They valued elegant solutions with closed-form expressions over large data models and questioned our drive for “accuracy.” For them, the models were simply tools to support sound decisions, not to spit out precise answers. To us, that was shortsighted—we had advanced technology and new ways to tackle complex problems.

Fast forward to 2024, and here we are again – new tools have emerged promising to reshape military decision-making. Artificial Intelligence, and more specifically, large language models (LLMs), may hold the potential to radically simplify how we approach the Military Decision-Making Process (MDMP) and Course of Action (COA) analysis. Unlike traditional tools like simulation and optimization, which rely on structured inputs and are often challenging to adapt in real-time, LLMs bring something new: the ability to engage conversationally, interpret complex contexts, and navigate the unstructured realities of multi-domain warfare in ways that classic models can’t. 

Or at least, that is the promise. In reality, we just don’t know yet. Will they replace other technologies, orchestrate them, or just tackle limited scopes based on the outputs of traditional approaches?

To understand where LLMs fit in, let’s first revisit the strengths and limitations of the methods we’ve relied on for decades—tabletop exercises, optimization, and simulation—and see how LLMs could add entirely new dimensions to COA analysis.

The Traditional Foundations: Tabletop Exercises, Optimization, and Simulation

In military planning, tabletop exercises (TTXs), mathematical optimization, and simulations are foundational tools. Each has distinct strengths and a well-established place in COA analysis and the MDMP.

1. Tabletop Exercises (TTXs): From Simple Rules to Realistic Modeling

Tabletop exercises have a long history in military planning dating back to the late 18th-century Prussian Kriegsspiel. TTXs began as map-based games designed to simulate battlefield scenarios and used rigid rules to help officers practice decision-making. As warfare grew more complex, TTXs evolved, incorporating closed-form mathematical expressions to reflect factors like troop movements, supply constraints, and timing—adding rigor and realism to exercises that had previously relied on straightforward, static assumptions.

Today, TTXs leverage optimization and simulation techniques to help commanders test responses and refine COAs in realistic, data-rich environments. Optimization supports resource allocation, while simulation introduces variability and risk, making TTXs more adaptable than before. However, TTXs are still limited by the static nature of pre-defined scenarios and heavy reliance on human judgment. They’re invaluable, but they struggle to deliver the agility and adaptability that modern multi-domain operations (MDO) demand.

2. Mathematical Optimization: Structured, Efficient Decision Support

Mathematical optimization emerged as a powerful tool for military planning, particularly during World War II, to maximize efficiency in logistics. Optimization’s strength lies in structured, resource-intensive scenarios where planners need to balance constraints like fuel, time, or personnel against operational objectives. It can weigh multiple COAs based on criteria like risk reduction, impact maximization, or resource minimization.

However, optimization models are ultimately restricted by the assumptions and constraints they require. They can struggle to adapt to unpredictable or rapidly changing battlefield conditions—an increasingly common requirement in MDO, where the environment may shift rapidly across domains. While optimization remains essential for planning in structured environments, it lacks the flexibility required for the real-time, nuanced insights needed on today’s battlefields.

3. Simulation: Bridging Theory and Reality

Simulation tools allow planners to test COAs in virtual environments that represent realistic battlefield dynamics. Simulations are highly effective for exploring multi-step strategies, logistics, and even high-stakes scenarios like nuclear responses. They can model probabilistic outcomes and complex interdependencies, providing commanders with a preview of how different actions might play out. Relative to optimization, they also have a much stronger ability to capture the stochastic nature of specific scenarios and they provide a better sense of variability and risk.

Yet simulations can be inherently resource-intensive and are typically constrained by pre-set parameters. They don’t easily adapt to emerging or unexpected information. Additionally, setting up and interpreting simulations often requires a high degree of technical skill, and they can’t always keep pace with real-time changes in high-pressure situations. This is where the unique capabilities of LLMs start to become apparent.

Where Do LLMs Fit? Dialogue, Interpretation, Orchestration, and Creation

LLMs bring capabilities to COA analysis and military planning that aren’t just supportive—they open up new frontiers in how we can approach complex, multi-domain operations. In addition to offering a conversational interface for real-time engagement, LLMs excel in interpreting and reasoning through complex, unstructured environments and strategic doctrines that traditional models struggle to encapsulate. But perhaps most remarkably, LLMs can go beyond orchestrating existing simulations and optimizations. They can create new models from scratch when needed, building analytical constructs that align precisely with immediate, real-world demands.

1. Real-Time Synthesis and Intuitive Interaction

One of the standout advantages of LLMs is their ease of use through conversational interfaces. Unlike simulation or optimization models, which require structured inputs and technical expertise, LLMs are designed to interact naturally with users in plain language. This capability gives LLMs a pronounced edge for time-sensitive environments without technical setup or a deep understanding of adjusting constraints or simulation conditions. It also gives less trained users, up and down the engagement space, an easy interface to engage and receive guidance and feedback.

Why This Matters: In a multi-domain environment, where users need immediate updates or clarity, LLMs can provide a direct line of communication, enabling rapid question-and-answer exchanges or updates through natural language. A commander might ask, “What potential vulnerabilities emerge if the weather shifts in our AO?” or “How would this COA align with coalition partner objectives in cyber and maritime domains?” Instead of requiring reconfiguration or recalibration, the LLM can draw on its vast data inputs and nuanced understanding to provide immediate, actionable answers. This real-time, conversational capability transforms accessibility, enabling faster, more fluid communication than traditional models can support.

2. Translating Complex Strategies and Doctrines into Real-Time Insight

Unlike optimization or simulation, which depend on rules and assumptions, LLMs excel in contexts that are more complex or abstract. They can interpret intricate doctrines, layered strategies, and context-specific nuances that are difficult to codify into traditional models. This flexibility is a powerful asset in COA analysis, where strategic objectives, cultural considerations, and multi-layered doctrines often resist rigid categorization.

Why This Matters: Multi-domain operations are rarely straightforward; they involve nuanced, layered strategies that interact across air, land, sea, space, and cyber domains. Translating these complex relationships into a model with fixed parameters can oversimplify or even misrepresent critical factors. But LLMs, with their capacity for language understanding and contextual reasoning, can bridge this gap.

For example, an LLM can interpret unstructured data, such as intelligence reports or doctrine, and distill it into insights aligned with operational goals. This capability means an LLM can handle complex COA scenarios that involve cultural factors, strategic doctrines, or coalition protocols—areas where rigid rules are less effective. An LLM could provide nuanced analysis, such as:

• Strategic Intent Interpretation: Describing how a given COA aligns (or conflicts) with overarching strategic objectives in language familiar to military planners, rather than as quantitative outcomes.

• Cultural and Environmental Sensitivity: Identifying region-specific cultural or local considerations that could impact mission success, translating soft intelligence into strategic insights that inform COAs.

• Doctrine Alignment: Contextualizing COAs in relation to strategic doctrines and policies, making it clear how different options uphold or deviate from established principles.

In short, LLMs can bridge the gap between structured analysis and the unstructured realities of modern operations, offering a new form of flexibility that complements the structure of traditional models.

3. Orchestrating Dynamic, Multi-Agent Feedback Loops

One of the most promising frontiers for LLMs in COA analysis is their potential for orchestrating dynamic operations across multiple agents. LLMs can serve as a coordination layer, facilitating real-time, feedback-driven COAs that adapt as conditions change. LLMs can orchestrate complex engagements through automation, situational awareness, and adaptive resource allocation by generating iterative, real-time updates and automatically distributing them to key agents.

Why This Matters:
Multi-agent orchestration is crucial in MDO, where success often depends on coordinated actions across multiple domains. LLMs can create feedback loops that continuously refine strategies based on the latest intelligence, integrating this information to support the Observe, Orient, Decide, Act (OODA) cycle. Imagine a scenario where an LLM continuously monitors changes in troop locations, environmental conditions, and enemy movements. As these conditions evolve, the LLM provides adaptive, coordinated instructions to each agent involved in the operation, ensuring that all parts of the engagement respond fluidly.

By automating responses and relaying real-time adjustments, LLMs can add coherence to operations that might otherwise be disconnected, empowering commanders to execute complex, fast-paced COAs without constant manual recalibration. This role as an orchestrator distinguishes LLMs, positioning them as tools that don’t just provide insights but actively manage dynamic, interconnected COAs across an operational landscape.

4. Everything Everywhere All at Once: LLMs as Dynamic Model Generators

Honestly, it kind of breaks my brain. LLMs are not only powerful enough to orchestrate simulations and optimizations, but they also have the capacity to reinvent these constructs on-the-fly. Imagine this: while we’ve traditionally relied on pre-defined models and frameworks, an LLM can dynamically generate entirely new models to address emerging, specific scenarios—quickly, autonomously, and accurately. This goes beyond enhancing simulations and optimization; it’s a radical shift, introducing a layer of adaptability that allows us to respond to real-world challenges as they arise, with custom analytical constructs generated in real time.

Picture a scenario where conditions on the ground call for a unique logistical or resource allocation model—one that isn’t in any playbook but is urgently needed. An LLM could draft this model from scratch, leveraging inputs, constraints, and objectives to build a fresh analytical approach. While simulation and optimization have long served us well by working within existing frameworks, LLMs allow planners to define brand-new structures, fundamentally transforming our ability to adapt to unexpected conditions with unparalleled strategic flexibility.

For instance, here’s a dynamically generated example of how an LLM could produce a GAMS optimization model for military resource deployment between two allies—the United States and North Torbia:

GAMS Code Example for Military Planning Model Generated by an LLM

This example model dynamically allocates resources from military bases to operation regions while minimizing costs, factoring in both transport and deployment requirements.

Why This Matters:

It’s just kinda nuts. For planners, this means access to custom solutions on demand, bypassing the lengthy process of building models from scratch. This real-time model generation accelerates response times and enables adaptive decision-making, which is crucial in environments where conditions evolve quickly—such as multi-domain operations or coalition efforts. Furthermore, the integration of these models directly into decision workflows empowers commanders to make data-driven decisions immediately, with models specifically tailored to address their operational needs.

In essence, LLMs’ ability to generate, interpret, and execute dynamic models positions them not merely as supportive tools but as adaptable, self-directed problem solvers, ensuring that military operations can keep pace with the complexities of the modern battlefield. This adaptability not only complements but enhances the utility of traditional methods, empowering planners to respond with precision and agility across domains.

Conclusion: The Unique Role of LLMs in Military Decision-Making

Integrating LLMs into COA analysis and MDMP is just beginning, so a healthy dose of skepticism is warranted, but the potential is unmistakable.  Simulation and optimization remain the backbone of military planning, indispensable for structured analysis and deep scenario exploration. But LLMs add something groundbreaking: they don’t just enhance these existing tools—they fundamentally expand what’s possible, how we engage with them, orchestrate them,  and generate new models in real-time to meet immediate, complex needs.

LLMs are not just another tool in the arsenal; they represent a new way to access, understand, and apply strategic information in real time, bridging the gap between traditional models and the unstructured demands of modern operations. This conversational, interpretive, and orchestration capability makes LLMs uniquely suited to COA analysis in multi-domain operations, where complexity and adaptability are mission-critical.

The realization that LLMs can reinvent the constructs of simulation and optimization themselves breaks traditional boundaries, opening up unprecedented adaptability. This means LLMs are not only making military planning faster or more efficient; they are transforming how we think about solving dynamic, high-stakes problems. 

In the world of MDO, the question of where LLMs fit isn’t about replacing traditional tools but amplifying their strengths while providing new, adaptive layers of intelligence. By acting as a dynamic interface—one that can synthesize information, engage in real-time dialogue, and create solutions on-the-fly—LLMs could soon become an indispensable asset in military decision-making. This is more than a new tool in the arsenal; it’s a transformative leap toward decision-making agility, positioning commanders to respond with confidence, precision, and speed in an increasingly complex operational landscape.