Autoregressive models (ARMs) suffer from a sequentiality bottleneck during reasoning, requiring step‑by‑step generation and thus limiting throughput. Diffusion language models (DLMs) can produce many candidates in parallel, but often lack semantic coherence. To address both issues we introduce Collaborative Reasoning Asymmetric Learning (CoRAL), built on an Asymmetric Planner‑Executor (APE) architecture.\
The APE architecture lets a DLM planner generate $K$ diverse reasoning trajectories in parallel. A large ARM executor then evaluates these trajectories, selects the most promising one, and synthesizes the final answer. To train the planner toward goal‑directed reasoning we propose Reward‑Guided Denoising for Chain‑of‑Thought (RDC), which injects outcome‑based rewards into the diffusion process, steering the generated paths toward the desired solution.\
Extensive experiments on vision‑language benchmarks demonstrate that CoRAL markedly outperforms prior diffusion‑based approaches and achieves performance comparable to strong ARMs, delivering both superior efficiency and answer quality.\
Review