In this talk, we present TALOS, a Tactical Adaptive LLM Orchestration System used to adapt NPC behaviour in a real-time strategy game demo. Players must combat waves of specialised enemy units to defend a central objective in a warehouse environment. An LLM-driven director analyses the player’s actions and adapts traditional behaviour trees to adjust enemy strategies and tactics at run time. We explain how we achieve the necessary performance running an LLM during gameplay on Arm-based mobile CPUs, how we fine-tuned and quantized the Gemma models used, and how we evaluated quality and performance. Through architecture details, performance analysis, tactical-quality evaluation, and demonstration, attendees will gain practical lessons for building adaptive game AI under tight latency and hardware constraints.
By the end of the talk the audience will
1. Understand how to run LLMs within the latency, memory, and thermal constraints of on-device gaming hardware.
2. How to design a hybrid game-AI architecture that combines an LLM tactical director with conventional behaviour trees.
3. How to apply practical techniques such as quantisation, fine-tuning, compact inputs, and bounded outputs to integrate an LLM into a game.
4. How to evaluate adaptive game AI using tactical-quality assessments, gameplay analytics, and on-device performance measurements.
