Background and motivation
Complex social systems have long been studied using simplified mathematical models inspired by physics and network science. These models help explain collective phenomena such as consensus formation, polarization and information diffusion.
Large language models (LLMs) now enable richer simulations in which agents communicate in natural language, maintain internal state and interact through social networks. IO Factory is a simulation framework for studying such systems. Recent work has used statistical mechanics to describe collective behaviour in networks of LLM agents. This motivates our central question: how well can simple, interpretable models explain the dynamics of more complex LLM social simulations?
Project objective and methods
The student will investigate whether statistical physics can provide useful abstractions of collective dynamics in IO Factory. An initial baseline will reproduce selected aspects of an existing physics-inspired approach and compare its assumptions and predictions with simulation results.
The project will then focus on one extension, chosen with the supervisors. Possible directions include moving beyond binary opinions using multistate or continuous models; representing interactions through social-media-like feeds; or estimating effective interaction parameters from simulation data. Depending on the chosen direction, the student could study consensus, polarization or fragmentation, including possible transitions between these collective states.
The work combines literature review, numerical modelling, controlled simulation experiments and statistical analysis. Models will be evaluated against simulation data, including how well they generalize across network structures or interaction settings. Identifying where a simplified model breaks down is also a valuable outcome. The precise scope will reflect the student's interests, available computational resources and the state of IO Factory when the project begins.
Indicative milestones
Dates will be agreed at project start and adapted to the thesis schedule.
- Complete a focused literature review, define the research question and establish a reproducible baseline.
- Run initial IO Factory experiments and compare the baseline model with observed collective dynamics.
- Implement one selected extension and evaluate it using additional simulations and sensitivity analyses.
- Deliver the thesis, documented code and a reproducible account of the experiments and findings.
Courses relevant from the bachelor's degree:
Recommended courses for theoretical curriculum in the master's programme: