Institutional Repository of Vilnius University
Visuomenės simuliacijos naudojant didžiuosius kalbų modelius /
Abstract
dc:descriptionIntegration of large language models into agent-based social simulations enables more flexible and diverse agent behavior modeling, but also introduces additional uncertainty and reproducibility challanges. This thesis analyzes the trade-off between behavioral diversity, reproducibility and computational costs when applying decision-making strategies. An experimentas agent-based social simulation was developed in which agents make location selection decisions within simplified town. Nine experimentas conditions were compared: rule based policy, fully LLM-based policy, and hybrid policies with two modes and controllable amount of LLM input. Each contitions was excecuted 30 times using different seeds. Model used was“openai/gpt-4o-mini“ model accessed though the OpenRouter API. Results showed hat partial LLM integration increased the diversity of agent selections, but in some cases reduced welfare related metrics and altered reproducibility metrics. Findings also demonstrated that the gate mode of hybrid policy significantly reduces the number of LLM calls while maintainning behavioral diversity level similar to the mix mode. Results show that hybrid decision making strategies can provide prantical compromise between behavioral divercity, reproducibility and computational cost in agent-based simulations.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gaidelis, Tadas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308337227&prefLang=en_US