TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs

Researchers introduce TimeLens, a family of multimodal large language models optimized for video temporal grounding that outperforms existing open-source models and even surpasses proprietary models like GPT-5 and Gemini-2.5-Flash. The work addresses critical data quality issues in existing benchmarks and introduces improved training datasets and algorithmic design principles.

• →TimeLens establishes new state-of-the-art performance in video temporal grounding among open-source models and surpasses proprietary models like GPT-5 and Gemini-2.5-Flash.
• →The research exposes critical quality issues in existing video temporal grounding benchmarks and introduces TimeLens-Bench with re-annotated datasets.
• →TimeLens-100K provides a large-scale, high-quality training dataset created through automated re-annotation pipeline.
• →Key algorithmic innovations include interleaved textual encoding for time representation and thinking-free reinforcement learning with verifiable rewards.
• →All codes, data, and models will be released open-source to facilitate future research in video understanding.