Newest papers

Beyond the Sequentiality Bottleneck: A Hybrid Diffusion-Autoregressive Framework for Efficient Multimodal Reasoning

Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28584 · cited by 0 works (OpenAlex) · open-access PDF

French

Les modèles autorégressifs (ARMs) font face à un "bottleneck de séquentialité" dans le raisonnement, tandis que les modèles de langage de diffusion (DLMs) offrent un parallélisme mais manquent de cohérence sémantique. Pour aborder les deux limitations, nous proposons Collaborative Reasoning Asymmetric Learning (CoRAL), un cadre hybride via une architecture Asymmetric Planner-Executor (APE): un DLM Planner génère K diverses trajectoires de raisonnement en parallèle, tandis qu'un grand ARM Executor sélectionne le plan optimal et synthétise la réponse finale. Pour former efficacement le Planner, nous introduisons Reward-Guided Denoising for CoT (RDC), qui intègre des

Machine translation (Meta M2M100); the English original follows.

Abstract

Autoregressive models (ARMs) face a "Sequentiality Bottleneck" in reasoning, while diffusion language models (DLMs) offer parallelism but lack semantic coherence. To addresses both limitations, we propose Collaborative Reasoning Asymmetric Learning (CoRAL), a hybrid framework via an Asymmetric Planner-Executor (APE) architecture: a DLM Planner generates K diverse reasoning trajectories in parallel, while a large ARM Executor selects the optimal plan and synthesizes the final answer. To train the Planner effectively, we introduce Reward-Guided Denoising for CoT (RDC), which integrates outcome-based rewards into the diffusion process to steer it toward goal-directed reasoning. Experiments on vision-language benchmarks show that CoRAL significantly outperforms prior diffusion-based methods and achieves competitive performance against strong ARMs, achieving both superior efficiency and quality.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Artificial intelligence, Planner, Process (computing), Face (sociological concept), Plan (archaeology)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange