WiseR explores reinforcement learning in a text-embedding space. A controller proposes an embedding, and the environment compares it with the embedding of a target token. The intended sequence of targets comes from the text being processed.
The prototype uses GPT-2 tokenization and embeddings to define that space. Its focus is the interaction between a proposed representation, a target, and a reward; it is not a complete language-generation system.
The implementation is unfinished. The environment needs a consistent state transition and reward definition before the idea can be evaluated. The open question is whether movement in the embedding space gives a useful learning signal beyond predicting the target token directly.