Generalized Intelligence for Tactical Decision-Making: Large Language Model-Driven Dynamic Weapon Target Assignment
Johannes Autenrieb, Ole Ostermann
TL;DR
This work tackles dynamic weapon-target assignment (WTA) in complex, uncertain tactical environments by integrating a centralized, context-aware reasoning module based on large language models (LLMs) into the control-loop framework. By preserving the conventional assignment–guidance decomposition and using Proportional Navigation Guidance (PNG) for continuous engagement, the approach replaces manually tuned surrogate costs with contextual reasoning that interprets mission data such as threat level, asset priority, and geometry. The authors demonstrate, through a simulation with multiple interceptors, targets, and defended assets, that LLM-driven reasoning can yield consistent, adaptable allocations with limited reassignments and latency comparable to traditional optimization methods, enabling explainable and robust decision support in autonomous defense systems. The results suggest that generalized AI can complement or partially replace explicit optimization in tactical WTA, paving the way for hybrid architectures that combine determinism with context-aware reasoning for improved resilience in dynamic engagements.
Abstract
Modern aerospace defense systems increasingly rely on autonomous decision-making to coordinate large numbers of interceptors against multiple incoming threats. Conventional weapon-target assignment (WTA) algorithms, including mixed-integer programming and auction-based methods, show limitations in dynamic and uncertain tactical environments where human-like reasoning and adaptive prioritization are required. This paper introduces a large language model (LLM) driven WTA framework that integrates generalized intelligence into cooperative missile guidance. The proposed system formulates the tactical decision process as a reasoning problem, in which an LLM evaluates spatial and temporal relationships among interceptors, targets, and defended assets to generate real-time assignments. In contrast to classical optimization methods, the approach leverages contextual mission data such as threat direction, asset priority, and closing velocity to adapt dynamically and reduce assignment switching. A dedicated simulation environment supports both static and dynamic assignment modes. Results demonstrate improved consistency, adaptability, and mission-level prioritization, establishing a foundation for integrating generalized artificial intelligence into tactical guidance systems.