Hu et al (2024) A Survey on Large Language Model-Based Game Agents
This is the most comprehensive survey of academic papers on Large Language Models being used as video game Agents (NPCs).
As a survey paper, for a student it provides a super valuable jumping off point into the research up to the publishing date. Starting here its possible to become very well read in the area in a short time. Perfect for a time poor MSc dissertation.
As far as original contribution goes the authors create a generic model for LLM based Agents which anyone could use as a guide for their own implementation.
Be wary though, the field is new, provides industry ready demos for employment and super easy to get started with (the proverbial low hanging fruit). However the flip side is exciting fluff is easy to make. Its much harder to push things forward in a formal way.
Park et al (2023) Generative Agents: Interactive Simulacra of Human Behavior
The authors propose and implement a generative agent architecture using LLMs to simulate believable human-like behaviour in NPCs. They achieve this by building in a cognitive architecture that simulates memory, beliefs, reflections, and intent.
Of particular value to a student (though its a double edged sword) is the demo, there is a comprehensive demo implementation with complete source code on github.
The authors completed an experimental analysis by removing the various parts of the cognitive architecture and reviewing the ‘believability’ of the agents. In doing so they showed that 4 parts are necessary.