Daffodil International University
Faculty of Science and Information Technology => Game Design => MCT => Games Model. => Topic started by: S. M. Monowar Kayser on April 15, 2026, 12:47:50 AM
-
Before the current wave of foundation models, game models of players and non-player characters were usually narrow, local, and brittle: player typologies were inferred from telemetry, non-player characters relied on behavior trees or dialogue graphs, and social simulation rarely extended beyond scripted triggers. Post-2020 research has transformed this landscape by treating game entities as memory-bearing, language-grounded agents whose behavior emerges from large-scale generative reasoning rather than only pre-authored branches. Park et al. (2023) showed that generative agents equipped with memory, reflection, and planning can produce believable social routines and emergent interactions in a sandbox environment; Wang et al. (2023) extended that logic to embodied game play by introducing Voyager, which accumulates executable skills in Minecraft through an automatic curriculum and a growing library of code-based behaviors; and DeepMind's SIMA project reframed the problem again by training an instructable agent that generalizes across multiple 3D environments through natural-language commands rather than game-specific APIs (Matthey et al., 2024). These developments mark a genuine departure from traditional game AI, because the model is no longer merely a policy over fixed states but a higher-level representation of memory, language, intention, and transferable capability. Even so, the gap between believability and reliability remains substantial: LLM-driven agents still suffer from memory drift, incoherent long-term goals, cultural bias, latency costs, and unpredictable social behavior that is aesthetically impressive but operationally difficult to ship in commercial games. Research now needs stronger longitudinal evaluation of trust, replayability, fairness, and narrative consistency, especially in multiplayer and live-service contexts where unstable agent behavior can create design and moderation problems. The most promising direction is therefore hybrid agent modeling in which symbolic constraints, retrieval systems, and player telemetry are layered around generative models so that social richness does not come at the expense of governance, performance, or designer control (Park et al., 2023; Wang et al., 2023; Matthey et al., 2024).
References
1. Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST 2023).
2. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., & Anandkumar, A. (2023). Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291.
3. Matthey, L., et al. (2024). Scalable instructable multiworld agent (SIMA): A generalist AI agent for 3D virtual environments. Google DeepMind technical report.
S. M. Monowar Kayser
Lecturer, Department of Multimedia & Creative Technology (MCT)
Faculty of Science & Information Technology
Daffodil International University (DIU)
Daffodil Smart City, Savar, Dhaka, Bangladesh
Visit: https://monowarkayser.com/ (https://monowarkayser.com/)