Vinci2 debuts EgoServe, a proactive-egocentric AI benchmark

The interesting question about a camera you wear all day is not whether it can see what you see. It is whether it should say anything about what it sees, and when. A new paper from Gong Sitong and colleagues, accepted at ECCV 2026 and posted to arXiv, tries to make that timing question something you can actually measure.

The system, Vinci2, is built around a training-free, memory-augmented agent the authors call EgoMemo. It keeps three parallel views of what the wearer has been doing: multi-scale temporal summaries, a semantic knowledge graph, and visual embedding archives. The claim is that these three memory types together let the agent decide when a proactive intervention is warranted, rather than either waiting to be asked or firing off alerts every time something interesting appears in frame. The authors frame this as a context-dependent decision problem, which is a more honest framing than the usual 'always-on assistant' pitch.

The more consequential contribution is probably the benchmark that comes with it. EgoServe contains over 3,000 service instances, organized across 4 temporal memory horizons that range from immediate safety alerts to long-term habit coaching, spanning 10 service categories. Until now, the 'when should the assistant speak' problem has mostly lived in vendor demo videos. A shared eval set means competing approaches can be compared on the same footing, which is what usually turns a research theme into a research race.

The honest caveat is that the arXiv abstract does not enumerate the 10 categories or name each of the four horizons, and 3,000 instances is a modest slice of the continuous wear that these systems are meant to survive in the wild. The paper also does not say how ground truth for 'the assistant should have intervened here' was collected, which is where a benchmark like this either earns trust or quietly leaks it.

Still, if you build wearable AI, the useful thing here is not the specific model. It is that restraint now has a scoreboard.