StoryTR: Narrative-Centric Video Temporal Retrieval with Theory of Mind Reasoning

Current video moment retrieval excels at action-centric tasks but struggles with narrative content. Models can see $\mathit{\text{what}}\mathit{\text{is}}\mathit{\text{happening}}$ but fail to reason $\mathit{\text{why}}\mathit{\text{it}}\mathit{\text{matters}}$. This semantic gap stems from the lack of $\mathbf{\text{Theory}}\mathbf{\text{of}}\mathbf{\text{Mind}}\mathbf{\text{(ToM)}}$: the cognitive ability to infer implicit intentions, mental states, and narrative causality from surface-level observations. We introduce $\mathbf{\text{StoryTR}}$, the first video moment retrieval benchmark requiring ToM reasoning, comprising 8.1k samples from narrative short-form videos (shorts/reels). These videos present an ideal testbed. Their high information density encodes meaning through subtle multimodal cues. For instance, a glance paired with a sigh carries entirely different semantics than the glance alone. Yet multimodal perception alone is insufficient; ToM is required to decode that a character smiling'' may actually be concealing hostility.‘’ To teach models this reasoning capability, we propose an $\mathbf{\text{Agentic}}\mathbf{\text{Data}}\mathbf{\text{Pipeline}}$ that generates training data with explicit three-tier ToM chains (intent decoding, narrative reasoning, boundary localization). Experiments reveal the severity of the reasoning gap: Gemini-3.0-Pro achieves only 0.53 Avg IoU on StoryTR. However, our 7B $\mathbf{\text{Shorts-Moment}}$ model, trained on ToM-guided data, improves +15.1% relative IoU over baselines, demonstrating that $\mathit{\text{narrative}}\mathit{\text{reasoning}}\mathit{\text{capability}}\mathit{\text{matters}}\mathit{\text{more}}\mathit{\text{than}}\mathit{\text{parameter}}\mathit{\text{scale}}$.