StreamOPD：一種用於串流影片理解的時空線索門控後訓練配方

摘要

串流影片理解要求直接從正在展開的影片之因果觀測前綴產生回應。現有系統加入推論時記憶、檢索與壓縮，然而一個無訓練的滑動視窗基準已經能與之匹敵。因此，我們固定一個無記憶的近期視窗協議，並探討僅靠後訓練能達到何種程度。具有可驗證獎勵的強化學習並不適合此設定，因為它會鼓勵產生冗長的「先思考、後回答」內容；而同策略蒸餾（OPD）能在學生軌跡上提供密集的 token 層級教師監督，但僅在兩個模型都以思考模式訓練時才穩定。這些觀察促成了 StreamOPD，一種結合可驗證串流影片資料、思考模式 OPD 與指令模式部署的方法。它將 StreamingBench 從 77.9% 提升至 83.9%——與 9B 教師模型相差不到 0.3 個百分點——並在不改變推論設定的情況下，將排除幻覺檢測子任務（HLD）後的 OVO-Bench 提升了 9.1 個百分點。作為教師特權的延伸，時空線索閘門（ST-CueGate）將「有線索相對於無線索」的教師似然比彙整為群體相對反應評分，用以重新加權 OPD。它在 OVO-Bench（排除 HLD）上達到 71.9%，在 Video-MME 上達到 64.9%，並且是唯一在所有四個基準上均維持在基礎模型之上的變體。將教師替換為學生初始策略的凍結副本——即同策略自我蒸餾——可保留大部分這些收益，並將 HLD 提升至 57.0%，高於未訓練的學生與 9B 教師，因此棄權損失並非此方法所固有。我們為開源串流影片研究提供了透明且可重現的參考。

English

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.