LLaVA-Scissor: Token Compression with Semantic Connected Components for Video LLMs Paper โข 2506.21862 โข Published Jun 27, 2025 โข 36