Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to StreamOPD, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from 77.9% to 83.9%---within 0.3 points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by 9.1 points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches 71.9% on OVO-Bench (excluding HLD) and 64.9% on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to 57.0%, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.</p>\n","updatedAt":"2026-08-18T15:59:28.857Z","author":{"_id":"6478679d7b370854241b2ad8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6478679d7b370854241b2ad8/dBczWYYdfEt9tQcnVGhQk.jpeg","fullname":"xiangan","name":"xiangan","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":19,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.85338294506073},"editors":["xiangan"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/6478679d7b370854241b2ad8/dBczWYYdfEt9tQcnVGhQk.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.16320","authors":[{"_id":"6a83c55f675db694db8cd497","name":"Keming Wu","hidden":false},{"_id":"6a83c55f675db694db8cd498","name":"Baoyi Wang","hidden":false},{"_id":"6a83c55f675db694db8cd499","name":"Kaichen Zhang","hidden":false},{"_id":"6a83c55f675db694db8cd49a","user":{"_id":"6478679d7b370854241b2ad8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6478679d7b370854241b2ad8/dBczWYYdfEt9tQcnVGhQk.jpeg","isPro":false,"fullname":"xiangan","user":"xiangan","type":"user","name":"xiangan"},"name":"Xiang An","status":"claimed_verified","statusLastChangedAt":"2026-08-18T16:45:06.420Z","hidden":false},{"_id":"6a83c55f675db694db8cd49b","name":"Zuhao Yang","hidden":false},{"_id":"6a83c55f675db694db8cd49c","name":"Sudong Wang","hidden":false},{"_id":"6a83c55f675db694db8cd49d","name":"Haowei Zhu","hidden":false},{"_id":"6a83c55f675db694db8cd49e","name":"Tingxuan Huang","hidden":false},{"_id":"6a83c55f675db694db8cd49f","name":"Hongcheng Gao","hidden":false},{"_id":"6a83c55f675db694db8cd4a0","name":"Bin Wang","hidden":false}],"publishedAt":"2026-08-17T00:00:00.000Z","submittedOnDailyAt":"2026-08-18T00:00:00.000Z","title":"StreamOPD: A Post-Training Recipe with Spatio-Temporal Cue Gating for Streaming Video Understanding","submittedOnDailyBy":{"_id":"6478679d7b370854241b2ad8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6478679d7b370854241b2ad8/dBczWYYdfEt9tQcnVGhQk.jpeg","isPro":false,"fullname":"xiangan","user":"xiangan","type":"user","name":"xiangan"},"summary":"Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to StreamOPD, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from 77.9% to 83.9%---within 0.3 points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by 9.1 points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches 71.9% on OVO-Bench (excluding HLD) and 64.9% on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to 57.0%, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.","upvotes":0,"discussionId":"6a83c55f675db694db8cd4a1","projectPage":"https://unix-ai-lab.github.io/StreamOPD/","githubRepo":"https://github.com/UniX-AI-Lab/StreamOPD","githubRepoAddedBy":"user","ai_summary":"StreamOPD improves streaming video understanding via on-policy distillation with verifiable rewards and a spatio-temporal cue-gating mechanism, achieving near-teacher performance without inference-time memory.","ai_keywords":["StreamOPD","on-policy distillation","verifiable rewards","thinking mode","instruct mode","Spatio-Temporal CueGate","ST-CueGate","on-policy self-distillation","abstention loss"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":4},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"query":{}}">
StreamOPD: A Post-Training Recipe with Spatio-Temporal Cue Gating for Streaming Video Understanding
Abstract
StreamOPD improves streaming video understanding via on-policy distillation with verifiable rewards and a spatio-temporal cue-gating mechanism, achieving near-teacher performance without inference-time memory.
Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to StreamOPD, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from 77.9% to 83.9%---within 0.3 points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by 9.1 points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches 71.9% on OVO-Bench (excluding HLD) and 64.9% on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to 57.0%, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.
Community
Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to StreamOPD, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from 77.9% to 83.9%---within 0.3 points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by 9.1 points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches 71.9% on OVO-Bench (excluding HLD) and 64.9% on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to 57.0%, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.16320 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.16320 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.