TTRL brings together several text-agent training experiments.
The TextWorld setup distills features from a language model into a student encoder before training a policy. A structured dialogue setup uses a small simulator with discrete actions. A generative-chat setup combines decoder pretraining with reinforcement learning over generated text.
The purpose is to compare how a language representation can support an agent's decisions, and how the training signal changes when the output becomes less constrained.
These are separate experimental pipelines. Results in a fixed simulator do not establish the quality of open-ended conversation, and a reward intended to encourage fluent text is only one part of evaluating the generated response.