This experiment combines a language model, discrete latent codes, and a learned policy for dialogue. The proposed workflow includes an offline learning stage followed by policy optimization using model-generated experience.
The implementation is exploratory and depends on external models, dialogue data, and supporting training components. The documented workflow also leaves dataset preparation incomplete.
The question is whether a policy over the latent representation can support useful conversation. The current project is a way to investigate that question, not evidence that the approach produces a reliable chatbot.