StreamVibe: Proactive Video Agents for Evolving User Intent

StreamVibe introduces a benchmark and model for proactive video agents that infer implicit user intents from continuous observations and dynamically update interactions as intents evolve. The approach includes StreamVibe-Bench with 4.1K evaluation and 85K training samples, StreamVibe-VL as a streaming proactive model, and StreamVibe-Harness for maintaining user intents as persistent states via hierarchical memory.