According to a new arXiv abstract, reinforcement learning has substantially improved reasoning in autoregressive and discrete-diffusion language models. The same abstract identifies language models that generate continuous token embeddings as an emerging alternative, but notes that applying RL to them remains a challenge.
The proposed approach is to backpropagate verifier rewards through continuous flow language models. This is framed as a way to bring the benefits of RL to this newer class of models.
The source is limited to the abstract, which is cut off mid-sentence. It does not include experimental details, benchmark results, or comparisons with existing methods, so those specifics are not reported here.