ALIVE Adds Proactive Monitoring and Feedback to Video LLMs
TL;DR
- ALIVE extends streaming video LLMs with proactive monitoring, streaming task coordination, and feedback-seeking interaction.
- Training uses two custom datasets, ALIVE-SFT-288K and ALIVE-RL-3K, with contrastive SFT then RL via Behavior-Aligned Credit Assignment.
- The authors release FSI-Bench and report ALIVE scoring highest overall against compared models on it plus four public benchmarks.
ALIVE is a streaming video LLM trained to do three things such models usually don't: alert the user without being asked, coordinate work across streaming turns, and request feedback mid-task. The system appears in a new paper on arXiv by Jinzhao Li, Yuan Xu, Yin Li, Miao Liu and co-authors.
Training runs in two stages over two custom datasets, ALIVE-SFT-288K and ALIVE-RL-3K. Contrastive supervised fine-tuning "teaches when and how to respond," and a reinforcement-learning phase uses what the authors call Behavior-Aligned Credit Assignment, which "combines trajectory-level and local advantages to refine decisions, responses, and silence across multi-turn interactions."
The paper also introduces FSI-Bench, a benchmark aimed at feedback seeking and the task completion that follows. ALIVE "achieves the highest overall scores among the compared models on four public benchmarks and FSI-Bench," per the abstract. The abstract names no baselines and publishes no per-benchmark numbers.
Originally reported by paper
Read the original article →Original headline: ALIVE Trains Streaming Video Model to Proactively Monitor and Seek Feedback Without Being Asked