Researchers introduced StreamVibe-Bench, a benchmark for user intent interaction. It contains 4.1K evaluation samples and 85K training samples. They also built StreamVibe-VL, a model that infers implicit user intent from continuous observations. StreamVibe-VL improves response timing and answer quality by 10.7% over baselines. Existing proactive models struggle with implicit and dynamic intents. StreamVibe-Harness maintains user intents as persistent states via hierarchical memory. It further improves response consistency and adaptation by 12.3%. The work targets proactive video agents for evolving user intent.

• 4.1K evaluation samples
• 85K training samples
• 10.7% better timing and answer quality
• 12.3% gain in proactive interaction