Making NPCs Think: Evaluating Post-Thinking for Real-Time Game NPC Dialogue Generation
Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 414-423
DOI: 10.1609/aiide.v22i1.43599 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, online erschienen am 08.11.2026. Verfasst von Hexi Wang und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Non-Playable Characters (NPCs) powered by large language models, commonly known as AI NPCs often struggle to create immersive and believable experiences for players, as traditional no-thinking models are prone to character drift across long conversations. Although reasoning models can resist this drift, their chain-of-thought reasoning before every response introduces latency that breaks immersion in real-time game play. Post-thinking, our recently introduced generation paradigm proposes a solution to improve a model’s character consistency with minimal perceivable latency by generating a reflection trace after each model response, and applying a sliding window of the most recent traces back into the subsequent turns to guide the model's future responses. We synthetically annotated NPC-RP, a 2,000 conversation dataset with three variants (no-thinking, pre-thinking, and post-thinking), and used it to fine-tune three Gemma3-4B-it models. We then evaluated the models through an automated latency benchmark, a trace tag compliance study, a blind expert survey on 10 characters, and a blind unity game evaluation. Through these experiments we have determined that post-thinking can effectively match pre-thinking on raw dialogue quality while adding nearly no perceivable latency over no-thinking in most scenarios. Post-thinking is also shown to be generally preferred over the no-thinking baseline in the game demo, and excels most at remembering and building on earlier conversations. These results demonstrate post-thinking's ability to deliver character consistency on par with reasoning models within the latency budget that real-time AI NPCs demand.
Zitieren
APA
Wang, H., & Carey, K. (2026). Making NPCs Think: Evaluating Post-Thinking for Real-Time Game NPC Dialogue Generation. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 414-423. https://doi.org/10.1609/aiide.v22i1.43599
Vancouver
Wang H, Carey K. Making NPCs Think: Evaluating Post-Thinking for Real-Time Game NPC Dialogue Generation. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):414-423. doi:10.1609/aiide.v22i1.43599
BibTeX
@article{Wang2026,
author = {Wang, Hexi and Carey, Keegan},
title = {Making NPCs Think: Evaluating Post-Thinking for Real-Time Game NPC Dialogue Generation},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
year = {2026},
volume = {22},
number = {1},
pages = {414--423},
doi = {10.1609/aiide.v22i1.43599}
}
EUVIMED