CoRAL pairs a diffusion planner with an autoregressive executor
Original title: Beyond the Sequentiality Bottleneck: A Hybrid Diffusion-Autoregressive Framework for Efficient Multimodal Reasoning
A diffusion planner drafts K parallel reasoning paths; a large autoregressive model selects one and writes the answer.
The paper proposes Collaborative Reasoning Asymmetric Learning, a hybrid of diffusion and autoregressive models for multimodal reasoning. A diffusion planner generates K diverse reasoning trajectories in parallel, and a large autoregressive executor chooses a plan and synthesizes the answer. Reward-Guided Denoising for chain-of-thought adds outcome-based rewards during diffusion training. On vision-language benchmarks, the authors report better results than prior diffusion methods and competitive results against strong autoregressive models.
Unconfirmed: Benchmark names, model sizes and numeric scores are not given in the abstract.
- The architecture is called Asymmetric Planner-Executor.
- Reward-Guided Denoising for CoT steers diffusion toward goal-directed reasoning.
- The paper claims gains in both efficiency and answer quality, without listing benchmark scores here.
- Products
- CoRAL
From the original article
1 min read at arxiv.orgAutoregressive models (ARMs) face a "Sequentiality Bottleneck" in reasoning, while diffusion language models (DLMs) offer parallelism but lack semantic coherence.
By Wei Cai · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.