DynFrame: Adaptive Reasoning for Video Understanding

A new framework called DynFrame addresses two structural gaps in video multimodal large language models (MLLMs). First, existing methods use a fixed per-window frame rate, forcing repeated retrieval calls for fine-grained evidence. DynFrame makes sampling density a learnable decision, emitting both the temporal window and sampling density as native tokens. Second, retrieval and answer generation are typically optimized with a single trajectory-level advantage, conflating credit for correct and incorrect steps. DynFrame decouples these, enabling more precise optimization. The framework is detailed in arXiv paper 2605.26680.