Video Large Language Models (VideoLLMs) excel in video understanding but suffer from significant computational overheads due to extensive visual tokens. Existing mainstream solutions including visual token compression and query-based Retrieval-Augmented Generation (Video RAG), face notable limitations: the former typically pursues task-agnostic compression of global video semantics, lacking query awareness, while the latter fails to optimize visual token construction and relies on costly, rigid annotation data. To address these challenges, we propose MERTA (Multi-Expert Retrieval for Token Allocation), a plug-and-play framework that integrates selective multi-expert collaborative retrieval with query-aware token allocation. MERTA selectively leverages lightweight experts to capture multimodal semantics, adaptively identifying key frames by measuring the relevance among queries, visual content, and auxiliary annotations. It then dynamically allocates the visual token budget, assigning higher budgets to key frames while keeping the total visual token budget constant. MERTA integrates seamlessly without requiring architectural modifications, demonstrating superior efficiency and performance over existing methods across multiple video understanding benchmarks.

Hybrid post-training combining supervised fine-tuning (SFT) and reinforcement learning (RL) is the standard paradigm for aligning large language models, yet fixed mixing schedules cannot adapt when the relative noise of the two signals evolves. We derive a noise-aware mixing weight by minimizing an MSE upper bound on the mixed stochastic gradient, yielding a closed-form that balances gradient noise variance and SFT–RL disagreement. Building on the token-wise reweighting of CHORD (CITATION)