TL;DR Can an AI watch a video stream and help before you ask? This paper tackles implicit and changing user intent with a benchmark, a model and an agent harness. Title: StreamVibe: Proactive Video Agents for Evolving User Intent URL: Key points ・🎯 Intent split into three levels: explicit, implicit, dynamic ・📚 StreamVibe-Bench: 4,100 samples, 20,546 help points, 183.4 hours of video ・🧠 StreamVibe-VL: adaptive SFT, then GRPO with trajectory-level rewards ・🔇 RL cuts false triggers: explicit FPR 2.02 vs. 12.75 for ViSpeak ・🗂 The Harness tracks active intents and progress, helping most on dynamic intents ・📈 Dynamic Pointwise F1 rises from 58.70 (SFT) to 66.34 (RL) ・⚠️ Long interactions remain hard: dynamic Trajectory C-F1 is only 8.91 The hard part of proactivity is knowing when to stay quiet, not just when to speak. #VideoAgents# #VLM#

Real-time video agents are here. Today, we’re sharing how we built Runway Characters, allowing you to turn one image into a fully expressive, conversational video agent streaming at 24 frames per second in HD. With just 1.75 seconds of end-to-end latency. Learn more below.

Agents, agents, agents at Ship London 🇬🇧 ▪︎ Video agents with @GoogleDeepMind ▪︎ Real-time voice agents with @ElevenLabs ▪︎ Agents in production with @Telegraph, @currys, @AKQA, and @easyJet June 17 at Magazine. Get your ticket.