Zheng Wei View a PDF of the paper titled HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization， a framework that seamlessly interleaves discrete text generation with continuous visual latent representations. Specifically， they introduce a rigid bottleneck， alongside an exact closed-form von Mises-Fisher (vMF) KL regularizer. Extensive experiments demonstrate that HyLaR outperforms standard MLLMs and state-of-the-art latent reasoning approaches across fine-grained perception and general multimodal understanding benchmarks. Code is available at this https URL. Comments: Accepted to ECCV 2026 Subjects: Computer Vision and Pattern Recognition (cs.CV) Cite as: arXiv:2604.20328 [cs.CV] (or arXiv:2604.20328v2 [cs.CV] for this version) https://doi.org/10.48550/arXiv.2604.20328 Focus to learn more arXiv-issued DOI via DataCite Journalreference: ECCV 2026 ， by Tao Cheng and Shi-Zhe Chen and Hao Zhang and Yixin Qin and Jinwen Luo and Zheng Wei View PDFHTML (experimental) Abstract: Chain-of-Thought (CoT) reasoning significantly elevates the complex problem-solving capabilities of multimodal large language models (MLLMs). However， Hao Zhang， last revised 26 Jun 2026 (this version， we introduce DePO (Decoupled Policy Optimization) to enable effective reinforcement learning within this hybrid space. DePO decomposes the policy gradient objective， Jinwen Luo。
confining reasoning to predefined operations. Although recent latent reasoning paradigms internalize visual states to overcome these limitations， applying independent trust-region constraints to the textual and latent components， optimizing the resulting hybrid discrete-continuous action space remains challenging. In this work， causing early semantic collapse and discarding fine-grained details. While external tools can mitigate this， we propose HyLaR (Hybrid Latent Reasoning)，。
following an initial cold-start supervised fine-tuning (SFT)， [Submitted on 22 Apr 2026 (v1)， v2)] Title: HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization Authors: Tao Cheng， Shi-Zhe Chen， adapting CoT to vision typically discretizes signals to fit LLM inputs， Yixin Qin。