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

Understanding a video requires more than recognizing isolated moments, as humans continuously track entities, states, and events over time. This capacity for visual state tracking is fundamental to video understanding, yet remains underexplored in current evaluations of Multimodal Large Language Models (MLLMs). We introduce Visual STAte Tracking benchmark (VSTAT), a video-based benchmark designed to diagnose visual state tracking in MLLMs. VSTAT consists of 834 clips drawn from both synthetic and real-world videos, paired with 1,500 questions that cannot be answered from any single frame or short segment, requiring continuous perception and integration of events across the entire video stream. Despite their strong performance on existing video benchmarks, we find that state-of-the-art MLLMs perform far below humans and only modestly above answer-prior baselines. To analyze this gap, we compare MLLMs' thinking traces with the underlying video stream to understand why and when MLLMs fail on VSTAT. We find that MLLMs reason and track correctly in text, but fail at visually perceiving the events they need to track. Finally, our preliminary evaluation suggests that recent agentic approaches, including MLLM-based video agents and coding agents, do not readily resolve these failures, still falling short on VSTAT.

Progress in AI has largely been driven by methods that assume less. As compute and data increase, approaches with weaker inductive biases generally outperform those with stronger assumptions. This is particularly characteristic of the field of Visual Representation Learning, where approaches have gone from being dominated by Supervised Learning, to Weakly Supervised Learning, to the now widespread success of Self-Supervised Learning without human labels. Yet, even modern Self-Supervised Learning approaches still depend on strong inductive biases such as augmentations, masking, or cropping. If this trend holds, even these remaining biases should become bottlenecks at scale -- and our experiments confirm this: the optimal strength of inductive biases decreases as data grows. This motivates the search for approaches that rely on fewer assumptions. To this end, we introduce Temporal Difference in Vision (TDV), a new paradigm for self-supervised learning from video that avoids existing inductive biases, relying instead on a causal assumption that the past causes the future. TDV functions by jointly training an image encoder and a motion encoder so that the current frame's representation plus the encoded motion equals the next frame's representation. Despite not leveraging any strong inductive biases, TDV matches state-of-the-art recipes on dense spatial tasks, laying the foundation for representation learning without strong assumptions.

Learning robust contextual knowledge from unlabeled videos is essential for advancing self-supervised tracking. However, conventional self-supervised trackers lack effective context modeling, while existing context association methods based on non-semantic queries struggle to adapt to unlabeled tracking scenarios, making it difficult to learn reliable contextual cues. In this work, we propose a novel self-supervised tracking framework, named \textbf{\tracker}, which introduces a dual-modal context association mechanism that jointly leverages fine-grained semantic prompts and contextual noise to drive the model toward learning robust tracking representations. Adherent to the easy-to-hard learning principle, our contextual association mechanism operates based on two stages. During early training, instance patch tokens (prompts) are assigned to both forward and backward tracking branches to facilitate the acquisition of tracking knowledge. As training progresses, contextual noise is gradually injected into the model to perturb feature, encouraging the tracker to learn robust tracking representations in a more complex feature space. Thus, this novel contextual association mechanism enables our self-supervised model to learn high-quality tracking representations from unlabeled videos, while being applied exclusively during training to preserve efficient inference. Extensive experiments demonstrate the superiority of our method.