NPC Brain explores how a game character's traits, current state, and history can affect its response to a situation. The work includes reactions, action choices, speech intent, and changes in state such as trust or fear.
Approach
A teacher model labels situations using an explicit character profile. Smaller student models learn distributions over the available actions and reactions. Profiles define the vocabulary and inputs so that the labeling process and runtime decision system can use the same meanings.
The intended game integration uses a compact student model inside Unreal through ONNX and NNE. The model would contribute strategic choices to an existing scored character controller. The development server is for experiments and labeling; free-form dialogue is a separate question.
Useful behavior depends on more than selecting a plausible action in isolation. Reactions need to remain consistent with the character's traits and remembered interactions, while changing when the situation changes. The experiments therefore include reviewed character responses and multi-turn behavior, rather than treating a single agreeable answer as success.