A new study finds that diffusion language models reason better when constrained to standard left-to-right generation. By avoiding arbitrary flexibility and using a simple training method called JustGRPO, researchers show that fewer options can expand reasoning capability rather than limit it.