Existing multimodal large language models for long-video understanding predominantly rely on uniform sampling and single-turn inference, limiting their ability to identify sparse yet critical evidence amid extensive redundancy. We introduce Video-o3, a novel framework that supports iterative discovery of salient visual clues, fine-grained inspection of key segments, and adaptive termination once sufficient evidence is acquired. Technically, we address two core challenges in interleaved tool invocation. First, to mitigate attention dispersion induced by the heterogeneity of reasoning and tool-calling, we propose Task-Decoupled Attention Masking, which isolates per-step concentration while preserving shared global context. Second, to control context length growth in multi-turn interactions, we introduce a Verifiable Trajectory-Guided Reward that balances exploration coverage with reasoning efficiency. To support training at scale, we further develop a data synthesis pipeline and construct Seeker-173K, comprising 173K high-quality tool-interaction trajectories for effective supervised and reinforcement learning. Extensive experiments show that Video-o3 substantially outperforms state-of-the-art methods, achieving 72.1% accuracy on MLVU and 46.5% on Video-Holmes. These results demonstrate Video-o3's strong multi-hop evidence-seeking and reasoning capabilities, and validate the effectiveness of native tool invocation in long-video scenarios.

Current Multimodal Large Language Models (MLLMs) struggle with long videos because they typically rely on uniform frame sampling and single-turn inference. This approach often dilutes critical visual evidence within redundant background content.

Video-o3 introduces a paradigm shift by mimicking human behavior. Instead of watching a video passively, it actively explores the content. The model iteratively discovers salient visual clues, inspects key segments with fine-grained detail, and adaptively terminates the search once sufficient evidence is acquired.

Key Features:

• Goal-Driven Exploration: Unlike models that scan the whole video coarsely, Video-o3 starts with a coarse scan and iteratively focuses on informative segments.
• Native Interleaved Tool Use: The model supports "clue seeking" and "answer reasoning" within a single shared context, rather than decoupled modules.

Video-o3 is designed to solve the challenges of attention dispersion and contextual efficiency in long-video processing. The framework operates on a Think-and-Tool cycle. The model generates structured directives containing temporal windows and visual token quotas. It dynamically invokes the VideoCrop tool to inspect target segments with adaptive spatiotemporal resolution.

Core Technical Innovations:

• Task-Decoupled Attention Masking: To prevent attention dispersion, TDAM isolates per-step concentration. During clue seeking, the model attends only to the global context; during answering, it focuses on high-resolution tool observations.
• Verifiable Trajectory-Guided Reward: To control context length and cost, we introduce a reward mechanism that balances exploration coverage with reasoning efficiency, encouraging the model to terminate precisely when evidence is sufficient.

Training a model to perform native interleaved tool invocation requires high-quality exploration trajectories, which are scarce in existing datasets. To address this, we developed a scalable automated data synthesis pipeline.

About Seeker-173K:

• Structure: The dataset is stratified into a four-quadrant taxonomy based on evidence cardinality and visual saliency, covering tasks from "Single-Clue Direct Answering" to complex "Multi-Clue Tool Invocation".
• Quality: Human verification is enforced through random sampling in all stages. The pipeline rigorously filters out flawed instances, preserving only those with sound logic and factual visual evidence.