ALIVE Adds Monitoring and Feedback to Streaming Video AI
The model was trained in two stages and evaluated on FSI-Bench, a measure of feedback-seeking and task completion.

ALIVE is a streaming video AI model built to monitor events and coordinate changing requests, adding feedback-seeking to systems that often handle user instructions one at a time.
The model combines ongoing event monitoring with the ability to manage multiple requests as instructions change. It can also ask users for clarification or additional input during a task.
ALIVE was trained through supervised fine-tuning followed by reinforcement learning. Its developers introduced FSI-Bench to evaluate how well a model seeks feedback and completes tasks. On FSI-Bench and four other public benchmarks, ALIVE posted the top scores.
The project is open source. Its browser demo can run locally and uses two GPUs for streaming speech recognition and synthesis alongside the model.