Teaching NPCs to Think Once: AI-Filled Lookup Tables for Believable Villagers

Game characters can feel alive if an AI decides what they do, but running an AI for every character, every moment, is slow and expensive. We test a cheaper idea: ask an AI once to fill in small tables of everyday psychology, then let every villager look up its choices in those tables for free. Each villager has a personality and beliefs about itself and others (“people mean well”, “I am overlooked”, “hard work pays off”). What happens to it slowly changes those beliefs, and the beliefs change how it acts. We built a test village and ran nine experiments. An independent AI rebuilt the whole engine from a written description alone, and its villages matched ours decision for decision. Different AIs filled the tables consistently enough to be useful, but each gave its village its own temperament. Villagers lived different lives in different runs, and the belief loop was what made them different. Changing one table row changed the village in a predictable way. The main conclusions held when we changed arbitrary settings. 100,000 villagers ran on an ordinary phone at about a quarter of a second per in-game hour. Memory, not speed, was the limit. In a first, small test rated by AI judges, the table answers were as believable as a live AI’s answers to the same moments. When an outside “mind” steered one villager, the villager did not become like the mind. It changed through how others treated it, and its relationships outlasted the control. Over 1,000 days with 1,000 villagers, villagers kept changing inside, but their relationships set and feuds slowly built up. Finally, a separate AI session that never saw our results rebuilt the engine and reran all nine experiments from written instructions alone. Every result that should match exactly did. The tables cost cents to fill once and nothing to run, so believable villagers could be cheap enough for small studios and phones. Supplementary files: a technical version of the paper, and the full engine, data, experiment scripts and an independent blinded replication (village-core-v1.0.1.zip). The same code and data are on GitHub: https://github.com/Alexander-Lithium/village-core Licences: paper and data under CC BY 4.0; source code under MIT. Contact: [email protected] Authorship verification: the author’s identity is committed to by two SHA-256 hashes. Each is the hash of a private file holding an identity statement and a random salt. Publishing a file later proves authorship without revealing anything now.Tier 1 (online identity): 27E0712EBFA638765935DFB339C2121758E9600F6E112FC79F575B141E13C07ATier 2 (legal identity): F4813463D2B380ED660787B32289C38F38C6336E988EF22664EB318670A5F306

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24211997
Primary Topic
Artificial Intelligence in Games
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Teaching NPCs to Think Once: AI-Filled Lookup Tables for Believable Villagers

Alexander Lithium
Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Games
preprint

Teaching NPCs to Think Once: AI-Filled Lookup Tables for Believable Villagers

Alexander Lithium
preprint en

Abstract

Game characters can feel alive if an AI decides what they do, but running an AI for every character, every moment, is slow and expensive. We test a cheaper idea: ask an AI once to fill in small tables of everyday psychology, then let every villager look up its choices in those tables for free. Each villager has a personality and beliefs about itself and others (“people mean well”, “I am overlooked”, “hard work pays off”). What happens to it slowly changes those beliefs, and the beliefs change how it acts. We built a test village and ran nine experiments. An independent AI rebuilt the whole engine from a written description alone, and its villages matched ours decision for decision. Different AIs filled the tables consistently enough to be useful, but each gave its village its own temperament. Villagers lived different lives in different runs, and the belief loop was what made them different. Changing one table row changed the village in a predictable way. The main conclusions held when we changed arbitrary settings. 100,000 villagers ran on an ordinary phone at about a quarter of a second per in-game hour. Memory, not speed, was the limit. In a first, small test rated by AI judges, the table answers were as believable as a live AI’s answers to the same moments. When an outside “mind” steered one villager, the villager did not become like the mind. It changed through how others treated it, and its relationships outlasted the control. Over 1,000 days with 1,000 villagers, villagers kept changing inside, but their relationships set and feuds slowly built up. Finally, a separate AI session that never saw our results rebuilt the engine and reran all nine experiments from written instructions alone. Every result that should match exactly did. The tables cost cents to fill once and nothing to run, so believable villagers could be cheap enough for small studios and phones. Supplementary files: a technical version of the paper, and the full engine, data, experiment scripts and an independent blinded replication (village-core-v1.0.1.zip). The same code and data are on GitHub: https://github.com/Alexander-Lithium/village-core Licences: paper and data under CC BY 4.0; source code under MIT. Contact: [email protected] Authorship verification: the author’s identity is committed to by two SHA-256 hashes. Each is the hash of a private file holding an identity statement and a random salt. Publishing a file later proves authorship without revealing anything now.Tier 1 (online identity): 27E0712EBFA638765935DFB339C2121758E9600F6E112FC79F575B141E13C07ATier 2 (legal identity): F4813463D2B380ED660787B32289C38F38C6336E988EF22664EB318670A5F306

Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Games
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon