The 20th annual Social Simulation Conference (SSC 2025) took place from 25th to 29th August 2025 at Delft University of Technology, Netherlands. The theme was “Social simulation in a socio-technical context: embracing societies’ complexities”.

Nick presented the following paper:

Malleson, N., & Manley, E. (2025, April 16). Towards the use of Foundation Models and Embeddings for Nuanced Perception and Decision-Making in Agents. Social Simulation Conference, Technical University of Delft, Netherlands.

Abstract: Following the emergence of powerful generative large language models (LLMs), there has been a flurry of interest in the use of LLMs to control agents in agent-based models. Proponents argue that using the information about humans and human behaviour contained within an LLM could lead to the creation of agents who exhibit more complex and nuanced behaviour than those whose actions are driven by traditional behavioural frameworks. This paper begins to explore the use of a specific concept that underpins LLMs; that of embeddings. An embedding is a vector-based numerical representation of a piece of text that captures aspects of its meaning and context. We hypothesise that conceptualising agents’ characteristics through embeddings, rather than with discrete state variables, may offer a more nuanced and expressive foundation for representing agent characteristics and behaviours. We demonstrate the potential of this approach by recreating the Schelling residential segregation model using rich text descriptions of household agents and converting these to embeddings as a means of defining agents. The results show how agents can self-organise into more diverse and emergent clusters than is possible when they are defined with a small number of discrete attributes. This offers a path toward more realistic, highdimensional representations of agent heterogeneity.

The paper is available here