` / `<tool_call>` / `<answer>` closures collapse under temperature sampling.\n**Tool Necessity Gap** — with a 64-frame overview, \"skip-tool\" becomes a shortcut and the GRPO advantage of calling vs. skipping flattens to zero.\n\nWe propose **PARA-GRPO** (Parseability-Anchored and Ratio-gAted GRPO), pairing one targeted fix per failure: (i) a format reward applied only at the structural-token positions most prone to collapse, and (ii) per-prompt overview-frame randomization K ∼ Uniform{4, 8, 16, 32, 64} that keeps the tool-call advantage non-degenerate.\n\nFully open: paper, code, weights, data\n📄 arxiv.org/abs/2605.20342 · 💻 github.com/EvolvingLMMs-Lab/ParaVT · 🤖 https://huggingface.co/ParaVT · 🌐 evolvinglmms-lab.github.io/ParaVT","html":"<p>Long-video understanding is becoming agentic where LMMs are post-trained with RL to natively invoke video tools (e.g., temporal cropping). But every existing native-RL recipe (including our own LongVT @ CVPR 2026) dispatches tool calls <em>sequentially</em>, one per turn: a bad crop has no peer correction, multi-turn calls drift the context, and inference cost grows linearly with turns.</p>\n<p><strong>ParaVT</strong> is the first multi-agent end-to-end RL-trained framework for <strong>Para</strong>llel <strong>V</strong>ideo <strong>T</strong>ool calling. A main agent emits multiple temporal-window crops in a <em>single</em> turn, weight-sharing sub-agents process them concurrently, and a gather-and-reason step produces the final answer.</p>\n<p>But applying standard GRPO on top of a tool-native LMM surfaces two coupled failures driven by the same pretrained tool prior. We call this the <strong>Tool Prior Paradox</strong>:</p>\n<p><strong>Format Fragility</strong> — SFT-learned <code>&lt;think&gt;</code> / <code>&lt;tool_call&gt;</code> / <code>&lt;answer&gt;</code> closures collapse under temperature sampling.<br><strong>Tool Necessity Gap</strong> — with a 64-frame overview, \"skip-tool\" becomes a shortcut and the GRPO advantage of calling vs. skipping flattens to zero.</p>\n<p>We propose <strong>PARA-GRPO</strong> (Parseability-Anchored and Ratio-gAted GRPO), pairing one targeted fix per failure: (i) a format reward applied only at the structural-token positions most prone to collapse, and (ii) per-prompt overview-frame randomization K ∼ Uniform{4, 8, 16, 32, 64} that keeps the tool-call advantage non-degenerate.</p>\n<p>Fully open: paper, code, weights, data<br>📄 arxiv.org/abs/2605.20342 · 💻 github.com/EvolvingLMMs-Lab/ParaVT · 🤖 <a href=\"https://huggingface.co/ParaVT\">https://huggingface.co/ParaVT</a> · 🌐 evolvinglmms-lab.github.io/ParaVT</p>\n","updatedAt":"2026-05-26T03:00:31.629Z","author":{"_id":"6524d665ab1416594149e07e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6524d665ab1416594149e07e/KMsCaAtV0DLC4tqN8f2a7.png","fullname":"Zuhao Yang","name":"mwxely","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":7,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7891132831573486},"editors":["mwxely"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/6524d665ab1416594149e07e/KMsCaAtV0DLC4tqN8f2a7.png"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.20342","authors":[{"_id":"6a0e9112164dbbc68a26c627","user":{"_id":"6524d665ab1416594149e07e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6524d665ab1416594149e07e/KMsCaAtV0DLC4tqN8f2a7.png","isPro":false,"fullname":"Zuhao Yang","user":"mwxely","type":"user","name":"mwxely"},"name":"Zuhao Yang","status":"claimed_verified","statusLastChangedAt":"2026-05-21T19:21:35.402Z","hidden":false},{"_id":"6a0e9112164dbbc68a26c628","user":{"_id":"64bb77e786e7fb5b8a317a43","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64bb77e786e7fb5b8a317a43/J0jOrlZJ9gazdYaeSH2Bo.png","isPro":false,"fullname":"kcz","user":"kcz358","type":"user","name":"kcz358"},"name":"Kaichen Zhang","status":"claimed_verified","statusLastChangedAt":"2026-05-26T07:52:10.658Z","hidden":false},{"_id":"6a0e9112164dbbc68a26c629","name":"Sudong Wang","hidden":false},{"_id":"6a0e9112164dbbc68a26c62a","name":"Keming Wu","hidden":false},{"_id":"6a0e9112164dbbc68a26c62b","user":{"_id":"654f99f74c8874c64d4e5664","avatarUrl":"/avatars/e9da0d688f91ae49db91d0ebebb3782a.svg","isPro":false,"fullname":"Zhongyu Yang","user":"yzzyu","type":"user","name":"yzzyu"},"name":"Zhongyu Yang","status":"claimed_verified","statusLastChangedAt":"2026-05-26T07:52:12.386Z","hidden":false},{"_id":"6a0e9112164dbbc68a26c62c","name":"Bo Li","hidden":false},{"_id":"6a0e9112164dbbc68a26c62d","name":"Xiaojuan Qi","hidden":false},{"_id":"6a0e9112164dbbc68a26c62e","name":"Shijian Lu","hidden":false},{"_id":"6a0e9112164dbbc68a26c62f","name":"Xingxuan Li","hidden":false},{"_id":"6a0e9112164dbbc68a26c630","name":"Lidong Bing","hidden":false}],"publishedAt":"2026-05-19T00:00:00.000Z","submittedOnDailyAt":"2026-05-26T00:00:00.000Z","title":"ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning","submittedOnDailyBy":{"_id":"6524d665ab1416594149e07e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6524d665ab1416594149e07e/KMsCaAtV0DLC4tqN8f2a7.png","isPro":false,"fullname":"Zuhao Yang","user":"mwxely","type":"user","name":"mwxely"},"summary":"Training large multimodal models (LMMs) via reinforcement learning (RL) to natively invoke video-processing tools (e.g., cropping) has become a promising route to long-video understanding. However, existing native-RL methods dispatch tool calls sequentially (i.e., one per turn): a single wrong crop propagates errors without peer correction, multi-turn tool calls corrupt context, and inference cost scales linearly with the number of turns. We introduce ParaVT, the first multi-agent end-to-end RL-trained framework for Parallel Video Tool calling, dispatching multiple time-window crops in a single turn for cleaner context and better fault tolerance. Yet applying standard RL to ParaVT reveals an obstacle we term the Tool Prior Paradox: the pretrained tool priors that enable tool exploration also destabilize cold-started structural format and expose the skip-tool reward shortcut under temperature sampling. A cross-model contrast on a weaker-prior LMM supports this claim: format stays stable but RL elicits zero tool calls, indicating that prior strength is the shared driver of both format collapse and tool exploration. We propose PARA-GRPO (Parseability-Anchored and Ratio-gAted GRPO), which augments standard RL with two complementary mechanisms: (i) a targeted format reward applied only at the structural-token positions most prone to collapse, and (ii) a per-prompt frame-budget randomization that creates training prompts where calling the tool yields a measurable reward signal over skipping it. Across six long-video understanding benchmarks, ParaVT improves over the Qwen3-VL baseline by +7.9% on average, with PARA-GRPO lifting training-time format compliance from 0.13 to 0.64. As tool capabilities become increasingly internalized in modern LMMs, RL must cooperate with the resulting priors, and ParaVT offers a general recipe for agentic RL. Code, data, and model weights are publicly available.","upvotes":27,"discussionId":"6a0e9113164dbbc68a26c631","projectPage":"https://evolvinglmms-lab.github.io/ParaVT/","githubRepo":"https://github.com/EvolvingLMMs-Lab/ParaVT","githubRepoAddedBy":"user","ai_summary":"ParaVT enables parallel video tool calling through multi-agent reinforcement learning, addressing limitations of sequential approaches and improving long-video understanding performance.","ai_keywords":["multimodal models","reinforcement learning","video-processing tools","parallel video tool calling","tool prior paradox","PARAGRPO","structural tokens","frame-budget randomization","long-video understanding","Qwen3-VL"],"githubStars":10,"organization":{"_id":"6583eb89bed3689928f5d845","name":"lmms-lab","fullname":"LMMs-Lab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/62d3f7d84b0933c48f3cdd9c/0sliNO9xGhOjVWw20A1Ge.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6524d665ab1416594149e07e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6524d665ab1416594149e07e/KMsCaAtV0DLC4tqN8f2a7.png","isPro":false,"fullname":"Zuhao Yang","user":"mwxely","type":"user"},{"_id":"66bf00ca5b4e241fe266059d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66bf00ca5b4e241fe266059d/VoWPC_C4zoeT6dS699t7L.png","isPro":false,"fullname":"Keming Wu","user":"wukeming11","type":"user"},{"_id":"649aa367c6cf3cc95bc1b7f6","avatarUrl":"/avatars/4bf5446c261eab08fc06caebf4c5779a.svg","isPro":false,"fullname":"Yifei Shen","user":"yshenaw","type":"user"},{"_id":"66100bacac50abb8d56dece6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66100bacac50abb8d56dece6/fd-4VMpb_1nl903yAIK4K.jpeg","isPro":false,"fullname":"Ding Yue","user":"dingyue1011","type":"user"},{"_id":"646a11791556443f24b582e9","avatarUrl":"/avatars/119ad9a19fd69448403fafa9ad8fcb6f.svg","isPro":false,"fullname":"Zonglin Yang","user":"ZonglinY","type":"user"},{"_id":"66d94fa1c81167fc5e50781a","avatarUrl":"/avatars/df61cce3e4298b5c3e0c4f81b7281b25.svg","isPro":false,"fullname":"Wenhao Li","user":"wenhaoli-xmu","type":"user"},{"_id":"64bb77e786e7fb5b8a317a43","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64bb77e786e7fb5b8a317a43/J0jOrlZJ9gazdYaeSH2Bo.png","isPro":false,"fullname":"kcz","user":"kcz358","type":"user"},{"_id":"6527b7280ae663e384eb8499","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6527b7280ae663e384eb8499/73yF3eu2cUx7jVZrhXnXx.jpeg","isPro":false,"fullname":"Senqiao Yang","user":"Senqiao","type":"user"},{"_id":"638f1803c67af472d317a922","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/638f1803c67af472d317a922/9BMVXqHa-AsdZPmBprcbd.jpeg","isPro":false,"fullname":"siyue zhang","user":"siyue","type":"user"},{"_id":"6690f58e2f9f6f9c88e91031","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6690f58e2f9f6f9c88e91031/QQ_VoEh7NlE6BUvii08zk.png","isPro":true,"fullname":"Sudong Wang","user":"xiao45791","type":"user"},{"_id":"655c70d331c4978366d4b2e6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/655c70d331c4978366d4b2e6/X-KjTNkxtzeYu9ngBOh_C.jpeg","isPro":false,"fullname":"yiyexy","user":"yiyexy","type":"user"},{"_id":"69d3fe4dc599864cfc5bddb5","avatarUrl":"/avatars/a720ed7e30f00414e6ed5590f989b0db.svg","isPro":false,"fullname":"s z","user":"simon77818","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6583eb89bed3689928f5d845","name":"lmms-lab","fullname":"LMMs-Lab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/62d3f7d84b0933c48f3cdd9c/0sliNO9xGhOjVWw20A1Ge.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.20342.md"}">

Get this paper in your agent:

Don't have the latest CLI?

Add this paper to a collection to link it from this page.