Simulacija osebnostnih lastnosti z velikimi jezikovnimi modeli

Savec, Borut (author), Robnik Šikonja, Marko (mentor)

Veliki jezikovni modeli omogočajo gradnjo generativnih agentov, ki na podlagi samoporočil posameznikov simulirajo njihove odgovore na osebnostne vprašalnike. Tuje raziskave so tak pristop že uspešno preizkusile v angleščini, v diplomski nalogi preverjamo, ali pristop deluje tudi v slovenščini. Z desetimi udeleženci smo izvedli polstrukturirane intervjuje in iz njih zgradili generativne agente. Ovrednotili smo korelacije med napovedmi odgovorov generativnih agentov in odgovori udeležencev na vprašalnik osebnostnih lastnosti BFI-44. Vsi agenti so dosegli zmerno stopnjo korelacije med 0,36 in 0,60, kar potrjuje, da pristop deluje tudi v slovenščini. Najuspešnejši je bil agent, zgrajen na modelu GPT4o. V nadaljnjih raziskavah bi bilo zanimivo ponoviti študijo z večjim vzorcem udeležencev, za isto osebo večkrat generirati agente na enakih podatkih ter eksperimentirati z različnimi parametri.

• Language: Slovenian
• Keywords: generativni agenti, veliki jezikovni modeli, simuliranje osebnostnih značilnosti
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-187777
• Publication date in RUL: 14.09.2026

Secondary language

• Language: English
• Title: Simulation of personality traits using large language models

Abstract

Large language models enable the construction of generative agents that simulate individuals’ responses on personality questionares based on their self-reports. Such aproach was already successfully tested in English, in this thesis, we examine whether the approach also works in Slovene. We conducted semi-structured interviews with ten participants and used these to build generative agents. We evaluated the correlations between the generative agents’ predicted responses and participants’ responses on the BFI-44 personality questionnaire. All agents achieved a moderate degree of correlation, between 0,36 and 0,60, confirming that the approach also works in Slovene. The agent built on GPT4o was the most successful. In further research, it would be useful to repeat the study with a larger sample of participants, generate agents on the same data for the same person multiple times, and experiment with different parameters.

• Keywords: Generative agents, Large language models, Simulation of personality traits