Gedas Bertasius

Our proposed VideoTreeSearch achieves state-of-the-art results across all grounded long-video QA benchmarks while also being more efficient than prior agentic video methods. We also show that we can achieve strong temporal grounding results even if we don't use any manually #VideoTreeSearch #LongVideoQA #AI #Nocountrycodesarementionedinthetweet.

We just released WatchAct, a benchmark for behavior-grounded robot manipulation (covered in the talk below). The robot has to watch a human video, infer what was done, make a plan, and then execute it. The best VLM (Gemini-3.1-Pro) only reaches 36.8% planning success rate, so #robotics #AI #machinelearning #Nocountrycodescanbeidentifiedintheprovidedtweet.

The gaps SVI-Bench reveals point to three research areas of broad current interest. First, video models with stronger reasoning, able to ground long-form analysis in observed visual evidence. Second, video generation models with explicit notions of multi-agent dynamics, capable #VideoModels #AIResearch #MultiAgentDynamics #Thetweetdoesnotmentionanyspecificcountries.Therefore #therearenocountrycodestoreturn.