DAMO: A Data-Efficient Multimodal Orchestrator for Temporal Reasoning with Video LLMS

Large Language Models (LLMs) have recently been extended to video, enabling video-grounded QA, dialogue, and temporal localization. Yet current Video LLMs struggle with fine-grained temporal reasoning, often failing to attribute answers to precise moments, particularly under limited supervision. We introduce DaMO, a data-efficient Video LLM explicitly designed for accurate temporal reasoning and multimodal understanding. At its core, the Temporal-aware Fuseformer adopts a hierarchical dual-stream design that progressively models modality-specific dynamics and fuses complementary visual–audio cues via lightweight learnable tokens. To improve efficiency without sacrificing semantics, a global residual reduces spatial redundancy while preserving global context. We further employ a four-stage progressive training regimen—video–text alignment, representation bridging, temporal perception, and dialogue tuning—augmented with LLM-generated temporally grounded QA pairs. Extensive experiments on temporal grounding and video QA benchmarks demonstrate that DaMO consistently surpasses prior methods, particularly on tasks requiring precise temporal alignment and reasoning, while using substantially less data and compute. Our project is released at https://yue-723.github.io/DaMO-2025-page/