Mielen toiminnan paljastaminen toiminnassa: Kognitiivisen päättelyn simulointi käyttäytymisessä suurten kielimallien persoona-agenttien avulla
Year: 2025
Type: dissertation
Abstract: Understanding human behaviour has wide-ranging applications across human-centred technologies, including robotics, educational systems, and autonomous vehicles. Behavioural understanding typically arises from two complementary perspectives and models: cognitive models representing the “why”, and behavioural models capturing the “what”. Despite the interconnectedness of both of perspectives, prior research has largely treated them in independently, which presents a critical gap. Cognitive models offer interpretive depth but are resource-intensive and static, while behavioural models are scalable yet often lack interpretability. Recent advancements in Large Language Model (LLM) persona agents offer a promising bridge between these approaches. However, there is a conundrum in the implementation of LLM persona agents, as existing implementations and research view these personas as either fictional or static, which fail to capture the true nature of dynamic and evolving real-world behaviours. To address these limitations, this thesis introduces a novel hybrid multimodal human-Artificial Intelligence (AI) workflow that integrates cognitive reasoning with visually grounded behavioural clusters, derived through Principal Component Analysis (PCA) and clustering techniques, resulting in adaptive LLM persona agents capable of simulating cognitive processes based on the behavioural data. The novel methodology is tested and validated using a large-scale dataset of over 100 million real-world interactions. Additionally, the findings respond to the critical gap of underdeveloped human-AI collaboration models in LLM persona research. The proposed methods can be adopted by practitioners and serves as a foundation for future research in dynamic, hybrid behavioural modelling with LLM persona agents.
Source: Aaltodoc (Aalto University)
Author: Tervola, Sonja
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7039556047