NTH

ALIVE: Making Streaming Video Models Interactive

AuthorsJinzhao Li, Yuan Xu, Jinxuan Wu, Yukun Wang, Yijia Lei, Hanyu Ji, Yinuo Chen, Zining Huang, Chi Zhang, Hang Ye, Yiwu Zhong, Honglei Yan, Panwang Pan, Yizhou Wang, Yin Li, Miao Liu

AffiliationsTsinghua University · ByteDance · Peking University · Beihang University · University of Wisconsin–Madison

November 20, 2026 2 min read
Watch on YouTube
The one-line take

ALIVE aims to make video-streaming AI an interactive partner that can monitor events, coordinate tasks, and ask for feedback.

Key results

288436
ALIVE-SFT-288K examples

Supervised fine-tuning interaction examples.

3523
ALIVE-RL-3K examples

Reinforcement-learning interaction examples.

40.3
IPIBench overall score

ALIVE’s score; 18.7 points above the strongest baseline.

18.7
IPIBench gain over strongest baseline

Point improvement over AURA.

52.8%
FSI-Bench EDP accuracy

Overall accuracy under the Explicit Decision Protocol.

54.0%
FSI-Bench FDP accuracy

Overall accuracy under the Free-form Decision Protocol.

What the paper found

ALIVE turns a streaming video model from a passive observer into an assistant that can monitor events, coordinate changing or concurrent requests, and ask users for clarification or actions when evidence is missing. Built on Qwen3-VL-8B-Instruct, it uses contrastive supervised fine-tuning to teach response timing and strategy selection, then Behavior-Aligned Credit Assignment, or BACA, to combine whole-trajectory rewards with local feedback on strategy, answers, and warranted silence. Training uses ALIVE-SFT-288K with 288436 examples and ALIVE-RL-3K with 3523 examples; the paper also introduces FSI-Bench for evaluating feedback seeking and task completion. ALIVE leads the compared models on OVO-Bench, StreamingBench, IPIBench, ProactiveVideoQA, and FSI-Bench. On IPIBench it scores 40.3, 18.7 points above the strongest baseline. On FSI-Bench, its accuracy is 52.8% under the Explicit Decision Protocol and 54.0% under the Free-form Decision Protocol, outperforming tested systems including OpenAI’s GPT-6 Astra and Google’s Gemini 3.5 Flash. The results show that post-training can teach a video model not just what to say, but when to act, wait, or seek user input.

Original abstract

Streaming video LLMs can continuously observe the world, yet often handle user requests in isolation. We introduce ALIVE, an interactive streaming video model supporting proactive monitoring, streaming task coordination, and feedback-seeking interaction. We construct ALIVE-SFT-288K and ALIVE-RL-3K for two-stage training. Contrastive supervised fine-tuning teaches when and how to respond, while reinforcement learning with Behavior-Aligned Credit Assignment (BACA) combines trajectory-level and local advantages to refine decisions, responses, and silence across multi-turn interactions. We also introduce FSI-Bench to evaluate feedback seeking and subsequent task completion. ALIVE achieves the highest overall scores among the compared models on four public benchmarks and FSI-Bench, demonstrating strong streaming understanding and interactive capabilities.

Read the original paper

More in Multimodal AI

Browse all 79 papers →