[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84388-en":3,"doc-seo-84388-105":29,"detail-sidebar-cat-0-en-105":83},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84388,7971461741311,"Ophelia","https://ap-avatar.wpscdn.com/avatar/74000253aff267980c6?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779345379180704826",8,"Research & Report","Switch-Reasoner: Learn When to Think in Multitask Mixtures via Reinforcement Learning","Multimodal Large Language Models (MLLMs) often follow a fixed Think-then-Answer paradigm, which is inefficient for heterogeneous multitask mixtures where some inputs need no explicit reasoning while others benefit from it. Learning when to think during post-training can also be unstable due to imbalanced rollouts, leading to always-thinking or always-direct behavior. Switch-Reasoner introduces a GRPO-based framework treating thinking as a virtual tool and adaptively selecting Direct vs Thinking Mode, stabilized by dual-level regulation and sample-level supervision.","arXiv :2607 .08572v 1 [ cs .CV] 9 Jul 2026  \nSWITCH-REASONER: LEARN WHEN TO THINK IN MULTITASK MIXTURES VIA REINFORCEMENT LEARNING  \nYiyang Fang 1 ,5 Pei Fu2 ∗ Jinjie Li3 Jian Liang 1 Wenke Huang4 Ruijie Luo 1 Shaojie Zhang2 Jian Luan2 Yi R. (May) Fung5 Mang Ye 1 ∗  \n1 Wuhan University 2 Xiaomi Inc 3 Wuhan University of Technology  \n4 Nanyang Technological University  \n5 The Hong Kong University of Science and Technology  \n[fangyiyang@whu.edu.cn](fangyiyang@whu.edu.cn) , [yemang@whu.edu.cn](yemang@whu.edu.cn)  \nABSTRACT  \nMultimodal Large Language Models (MLLMs) often follow a fixed Think-thenAnswer paradigm, which is inefficient in heterogeneous multitask settings because simple inputs may not require explicit reasoning while difficult ones can benefit substantially from it. Learning when to think is also unstable during post-training, where imbalanced rollouts can drive the model toward always-thinking or alwaysdirect behavior. We propose Switch-Reasoner, a GRPO-based framework that learns to adaptively select reasoning modes for MLLMs. It treats thinking as a virtual tool invocation and allows the model to either answer directly or invoke explicit reasoning before answering. To stabilize this decision, we introduce a dual-level regulation mechanism that balances the overall use of Thinking Mode and Direct Mode while providing sample-level supervision based on the relative benefit of the two choices. Experiments on 11 multimodal tasks show that SwitchReasoner reduces unnecessary reasoning while maintaining strong performance, achieving a better accuracy-efficiency trade-off. Code is available at [https:](https:)//[github.com/fuyyyyy/Switch-Reasoner](github.com/fuyyyyy/Switch-Reasoner).  \n1 INTRODUCTION  \nMultimodal Large Language Models (MLLMs) (Liu et al., 2023 ; Li et al., 2024a) have achieved remarkable progress in visual understanding and reasoning, supporting a broad range of tasks such as mathematical problem solving, scene understanding, chart reasoning, and visual question answering (Chen et al., 2024 ; Li et al., 2024b ; Wang et al., 2024 ; Ye et al., 2025) . As these models become increasingly general-purpose, reinforcement learning (RL)-based post-training (Liu et al., 2025a ; Zhou et al., 2025) has emerged as an effective approach for improving their reasoning capabilities beyond supervised fine-tuning (Schulman et al., 2017 ; Rafailov et al., 2023) . In particular, Group Relative Policy Optimization (GRPO) has become widely adopted for reasoning-oriented training (Guo et al., 2025 ; Shao et al., 2024 ; Ramesh et al., 2024) because of its simplicity and strong empirical performance without requiring a separately trained value model.  \nHowever, improving reasoning accuracy alone is insufficient for practical MLLM deployment. Existing RL-based reasoning pipelines typically encourage a fixed Think-then-Answer behavior (Yao et al., 2025), where the model generates an explicit chain of thought before producing its final answer. Although such deliberation is useful for difficult problems (Feng et al., 2025 ; Huang et al., 2025a ; Rong et al., 2025), it incurs substantial token, latency, and deployment costs. More importantly, this strategy assumes that all inputs require the same amount of reasoning, which is fundamentally mismatched with heterogeneous multimodal tasks (Wu et al., 2025) . As illustrated in Figure 1, some tasks benefit substantially from explicit reasoning, whereas others can often be solved reliably through direct responses. A desirable MLLM should therefore selectively switch between a Thinking Mode and a Direct Mode, invoking reasoning only when necessary.  \nRecent work has pursued reasoning efficiency from two main directions. Early-exit (Yang et al., 2025a) and reasoning-pruning (Hou et al., 2025) methods reduce redundant computation by shortening chain-of-thought generation after reasoning has already begun (Nagle et al., 2026) . While effective at  \n∗ Corresponding author. Work done by Yiyang Fan","cbCaiuDoJ99rHC0A","https://ap.wps.com/l/cbCaiuDoJ99rHC0A","pdf",2101302,1,19,"English","en",105,"# Abstract\n# Introduction\n## Motivation: Selective reasoning in multitask mixtures\n## Related work: reasoning efficiency and adaptive reasoning\n## Challenge under GRPO training","[{\"question\":\"How does Switch-Reasoner decide between Direct Mode and Thinking Mode?\",\"answer\":\"Switch-Reasoner treats thinking as a virtual tool invocation. It learns a policy that adaptively selects whether to answer directly or invoke explicit reasoning before answering, using GRPO and stabilization via dual-level regulation and sample-level supervision.\"}]",1784195240,48,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":78,"head_meta":80,"extra_data":82,"updated_unix":27},"switch-reasoner-learn-when-to-think-in-multitask-mixtures-via-reinforcement-learning","",{"@graph":35,"@context":77},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/switch-reasoner-learn-when-to-think-in-multitask-mixtures-via-reinforcement-learning/84388/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71],{"name":72,"@type":73,"acceptedAnswer":74},"How does Switch-Reasoner decide between Direct Mode and Thinking Mode?","Question",{"text":75,"@type":76},"Switch-Reasoner treats thinking as a virtual tool invocation. It learns a policy that adaptively selects whether to answer directly or invoke explicit reasoning before answering, using GRPO and stabilization via dual-level regulation and sample-level supervision.","Answer","https://schema.org",{"og:url":51,"og:type":79,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":81,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":84},[85,89,93,97,102,107,112,115,120,123,127],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":86,"show_sort_weight":87,"slug":88},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":90,"show_sort_weight":91,"slug":92},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Exam",70,"exam",{"id":98,"doc_module":4,"doc_module_name":45,"category_name":99,"show_sort_weight":100,"slug":101},5,"Comic",60,"comic",{"id":103,"doc_module":4,"doc_module_name":45,"category_name":104,"show_sort_weight":105,"slug":106},6,"Technology",50,"technology",{"id":108,"doc_module":4,"doc_module_name":45,"category_name":109,"show_sort_weight":110,"slug":111},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":113,"slug":114},30,"research-report",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":45,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":124,"slug":126},10,"Lifestyle","lifestyle",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":98,"slug":129},"General","general"]