Agenda

Non‑player characters (NPCs) are moving beyond pre‑scripted trees to responsive agents that converse, plan, and act in context. In this talk we share practical lessons from an Arm Developer Labs project run in partnership with North Carolina State University (NCSU), where a team of Master's students is building a proof‑of‑concept, puzzle‑style game that showcases LLM‑augmented NPCs running on Arm‑based devices.

We walk through the end‑to‑end approach the student team uses to keep gameplay smooth while adding intelligence: selecting and sizing models for on‑device inference, applying quantization to meet 30–60 FPS budgets, and using structured generation so model outputs map deterministically to in‑engine actions.

Takeaway

Game Design Insights: See how AI‑driven NPCs can enhance puzzle gameplay and player engagement.

Lessons Learned: Honest review of challenges, trade‑offs, and best practices from the student project.