Search papers, labs, and topics across Lattice.
To resolve the optimization conflicts that arise when jointly training for temporal grounding, general comprehension, and multi-step reasoning, the authors train domain-specific RL expert policies and consolidate them into an 8B student via Multi-Teacher On-Policy Distillation (MOPD). The framework supervises student-generated trajectories using routed teacher feedback paired with Reliability-Aware Informative Sampling (RAIS) to prioritize high-confidence supervision on instances with large teacher-student capability gaps. Video-MOPD-8B achieves open-weight state-of-the-art performance across video temporal grounding, general video QA, and multimodal STEM reasoning benchmarks at the 8B scale.
Domain-specialized RL experts can be unified without negative transfer by distilling their feedback directly onto student-generated trajectories, resolving the multi-task optimization bottleneck that limits single video foundation models.
Video understanding demands a convergence of complementary capabilities across perception, temporal understanding, and complex reasoning, which are difficult to jointly optimize within a single model. We introduce Video-MOPD-8B, an open-weight model dedicated to video understanding tasks. To fundamentally enhance its capabilities, we conduct targeted reinforcement learning (RL) optimization across three core domains: video temporal grounding (VTG), general video comprehension, and video STEM reasoning. We then unify their complementary capabilities via Multi-Teacher On-Policy Distillation (MOPD), which consolidates expert knowledge by supervising student-generated trajectories with routed teacher feedback. We further introduce Reliability-Aware Informative Sampling (RAIS), which selects examples with consistently reliable teacher supervision and large teacher-student performance gaps. Together, these components enable Video-MOPD-8B to achieve coordinated and comprehensive performance gains across diverse video understanding tasks. Extensive experiments on comprehensive benchmarks covering general video understanding, temporal grounding, video reasoning, and video STEM tasks demonstrate that Video-MOPD-8B achieves state-of-the-art performance among existing models at a comparable scale. The trained model weights are available at https://huggingface.co/LandH/Video-MOPD-8B.
