BoxTuning: Directly Injecting the Object Box for Multimodal Model Fine-Tuning

arXiv:2604.11136v1 Announce Type: cross Abstract: Object-level spatial-temporal understanding is essential for video question answering, yet existing multimodal large language models (MLLMs) encode frames holistically and lack explicit mechanisms for fine-grained object grounding. Recent work addresses this by serializing bounding box coordinates as text tokens, but this text-coordinate paradigm suffers from a fundamental modality mismatch: object information is inherently visual, yet encoding i