StreamOPD：面向流式视频理解的时空线索门控后训练方案

摘要

流式视频理解要求从正在展开的视频的按因果方式观察到的前缀中直接产生响应。现有系统增加了推理时记忆、检索和压缩，但一种免训练的滑动窗口基线已经能与它们匹敌。因此，我们固定一个无记忆的近期窗口协议，并探究仅靠后训练能达到何种程度。基于可验证奖励的强化学习并不适合这种场景，它会鼓励冗长的“先思考后回答”生成；而同策略蒸馏（OPD）能在学生轨迹上提供密集的词元级教师监督，但只有当教师和学生模型都以思考模式训练时才是稳定的。这些观察促成了 StreamOPD，这是一种结合可验证流式视频数据、思考模式 OPD 与指令模式部署的方案。它将 StreamingBench 从 77.9% 提升到 83.9%——与 9B 教师相差不到 0.3 个百分点——并在推理不变的情况下，将排除幻觉检测子任务（HLD）的 OVO-Bench 提升了 9.1 个百分点。作为教师特权扩展，时空线索门控（ST-CueGate）将“有线索”与“无线索”的教师似然比聚合成一个组相对响应分数，用于重新加权 OPD。它在 OVO-Bench（排除 HLD）上达到 71.9%，在 Video-MME 上达到 64.9%，并且是唯一在所有四个基准上保持在基础模型之上的变体。将教师替换为学生初始策略的冻结副本——即同策略自蒸馏——保留了大部分收益，并将 HLD 提高到 57.0%，高于未训练的学生模型和 9B 教师，因此弃权损失并非该方案所固有。我们为开源流式视频研究提供了一个透明且可复现的参考。

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.