AI 大模型精选文章

Temporal Self-Distillation: Learning Visual State Tracking in Videos Without Supervision

#S3T#Video-LLM#Self-Distillation#Visual State Tracking#VSTAT

本篇解读对应 Temporal Self-Distillation: Learning Visual State Tracking in Videos Without Supervision

arXiv。 arXiv:2609.04203(Submitted 3 Sep 2026,v1)

摘要

We introduce S3T (Self-Supervised Self-Distillation over Time), which, to the best of our knowledge, is the first fully self-contained framework for continuous video state tracking. Our method treats temporal sampling density as privileged information, based on the hypothesis that a denser view of the same clip recovers the running state more accurately. This view serves as the teacher, while a sparse-view student with the same weights learns to match its next-token distribution. The model generates its own target, so training requires no labels, separate teacher, or reward signal, and adds no inference cost. On LLaVA-OneVision-2-8B, S3T improves VSTAT accuracy by +1.74 as a single model, +2.38 with souping, and +2.70 with additional vision-encoder adaptation, while prior self-evolving methods leave state tracking largely unchanged. The capability learned from unlabeled synthetic clips transfers to real videos, improving performance by +7.95 on VSTAT-YouTube state-tracking questions and +4.50 on MVBench Action Count.

摘要给出方法名 S3T(Self-Supervised Self-Distillation over Time):把时间采样密度当作特权信息;密采样视图作 teacher,同一套权重的疏采样视图作 student,匹配 next-token 分布。监督来自模型自身输出,因此不需要标签、独立 teacher 或 reward,也不增加推理成本。在 LLaVA-OneVision-2-8B 上,VSTAT 精度报告为单模型 +1.74、souping +2.38、再加视觉编码器适配 +2.70;从无标注合成片段学到的能力迁移到真实视频时,VSTAT-YouTube 状态跟踪题为 +7.95,MVBench Action Count 为 +4.50。

总结

该工作把时间采样密度当作特权信息:同一段无标注片段的密视图作 teacher,同一冻结骨干加一个 LoRA adapter 的疏视图作 student,目标是模型自己的 next-token 分布。训练不使用标签、独立 teacher 网络、reward model 或 tracker,也不增加推理成本。VSTAT 上报告的提升为单模型 +1.74、souping +2.38、再加视觉编码器适配 +2.70,骨干为 LLaVA-OneVision-2-8B。迁移数字为 VSTAT-YouTube 状态跟踪题 +7.95、MVBench Action Count +4.50。同一目标在文中测试的其他骨干上并未一致迁移。

详细解读

VSTAT 所测量的差距

引言将困难定位于累积视觉状态:Video-LLM 须在物体被加入、移除、移动或遮挡时维持运行计数与场景状态。既有目标——无论是单答案监督微调还是视频–文本对比对齐——并不必然要求在整段片段上累积证据。

VSTAT [64] is a benchmark designed to test this cumulative state reasoning, and it shows that human performance reaches 90.5%, while the strongest open-source models score only 34–35%.

这两段区间属于引言中对 VSTAT 的引用。句子未点名那些开源模型,也不等于后文表格中的 34.74,更未声称 S3T 填平了与人类的差距。

把时间密度当作特权信息

The central idea is to use temporal sampling density as privileged information. We hypothesize that a denser sampling of the same sequence provides more evidence about the running scene state (such as object counts through occlusion) than a sparse sampling, while requiring no labels or annotations.

Teacher 是同一片段的更密采样,不是外部网络,也不是带标注答案。Student 匹配 teacher 的 next-token 分布。该假设针对运行中的场景状态,而非动作识别或时刻定位。这句话尚未固定两种帧数,也尚未指定后文表格所使用的骨干。

The student view samples k_s=12 frames and the teacher view samples k_t=24 frames over the same span, so the only difference is temporal density: both cover [0,T-1] with no crop, window, or sub-span.

两种视图使用同一冻结基座与同一个 LoRA adapter;差别在于帧密度以及是否保留梯度。原文并写明 student 的帧索引不是 teacher 的子集:两条均匀网格只共享端点。讨论部分随后报告,增益只出现在接近 12-frame student 与 24-frame teacher 的组合附近。这些帧数是训练时的构造。主表评测使用官方 64-frame 预算;原文并未把 k_s=12 等同于该预算。

在 LLaVA-OneVision-2-8B 上报告的提升

On LLaVA-OneVision-2-8B, S3T improves VSTAT accuracy by +1.74 as a single model, +2.38 with souping, and +2.70 with additional vision-encoder adaptation, while prior self-evolving methods leave state tracking largely unchanged.

三个增量对应同一骨干的三种配置,而不是三个数据集。引言把其中最强的一项复述为从 34.74 到 37.44(+2.70),并集中在 Count(+4.9)、Atomic(+5.1)、Sequence(+5.8)。与先前 self-evolving 方法的比较是原文在该基准上的读法,并不等于每一种 self-evolving 配方在每一类视频任务上都会失败。默认 adapter 设定冻结视觉编码器;+2.70 包含额外的视觉编码器适配。

The capability learned from unlabeled synthetic clips transfers to real videos, improving performance by +7.95 on VSTAT-YouTube state-tracking questions and +4.50 on MVBench Action Count.

训练片段是无标注 StateGen 视频。两个迁移数字是任务特定的:VSTAT-YouTube 状态跟踪题与 MVBench Action Count。原文并未报告在每一项真实视频基准上都有均匀提升,并写明一般视频理解性能得到保持,而不是每一项非计数任务都提高。

同一目标未一致迁移之处

S3T does not transfer consistently across the models we tested. We applied the same objective and data to eleven base models from five families and retuned the training settings for each model (App. L). Only LLaVA-OV-2-8B [2] showed a clear improvement.

这一限制由原文直接陈述。文中测试了十种候选解释,均未能说明该差异。文本中留下的两种可能性是:各骨干所接受的视频后训练量,以及低秩语言模型 adapter 能在多大程度上改变对视觉证据的使用。公开检查点无法把这两项因素分开。因此主表描述的是该目标下的 LLaVA-OneVision-2-8B / LLaVA-OV-2-8B,并不构成与骨干无关的 Video-LLM 通解。

证据边界

标题增量(VSTAT 上的 +1.74、+2.38、+2.70;VSTAT-YouTube 状态跟踪题 +7.95;MVBench Action Count +4.50)属于在 300 段无标注 StateGen 片段上训练的 LLaVA-OneVision-2-8B:训练时为 12-frame student 与 24-frame teacher,主比较使用官方 64-frame VSTAT 评测预算。同一目标与数据在 App. L 报告的另外十个基座上没有给出明确增益。该工作并未声称填平 90.5% 的人类 VSTAT 数字,未引入带标注 tracker,也不在适配权重之外增加推理成本。独立复现不属于本预印本。