The SkyrimNet mod integrates large language models (LLMs) into the game, allowing NPCs to generate dialogue based on environmental context and past experiences. Users reported that while the system initially provides high immersion, NPCs often become repetitive, fixating on specific traits from their pre-written biographies.
The author observed that NPCs like the necromancer Nymphanet consistently steer every conversation toward their primary profession. This phenomenon occurs because certain keywords in a biography act as 'semantic attractors,' forcing the LLM to prioritize specific associations over diverse topics.
Experiments with the character Fastred showed that attempting to restrict specific behaviors, such as fear, often backfires. The LLM compensated by expressing these thoughts in internal monologues or diaries, or by fixating on the forbidden topic even more intensely, similar to psychological suppression.
The author tested various hypotheses, including removing restrictive biography entries and introducing broader interests to prevent single-topic loops. While shifting the focus helped, the model often latched onto new 'anchors' like morality, leading to similarly repetitive behavior patterns.
The study suggests that LLMs treat dialogue as a task to fulfill based on provided biographies. When biographies are too detailed or contain strong thematic anchors, the model struggles to maintain conversational variety, often ignoring context in favor of its core programmed identity.
