Vinci2: Proactive Video Assistant Decides When to Interrupt Based on Continuous Egocentric Video

机器之心 · wechat · 2026-08-17

A joint team from Dalian University of Technology, AlayaLab, and the University of Tokyo introduces Vinci2, formalizing proactive assistance in first-person video assistants as a joint decision-and-generation task over continuous video streams. It comprises EgoServe, the first benchmark for proactive egocentric video assistance, and EgoMemo, a training-free memory-augmented agent. Accepted by ECCV 2026, code and data are open-sourced.

EgoServe covers over 128 hours of video and 3,400+ service instances, organized into four levels (immediate, short-term, situational, long-term) and ten subcategories, evaluating both temporal precision and response quality.

EgoMemo incrementally builds multi-scale temporal memory, an evolving knowledge graph, and visual embedding archives, using triple parallel retrieval and VLM description reconstruction for proactive intervention. It significantly outperforms Qwen3-VL-Plus and GPT-5-mini on EgoServe, especially in situational and long-term services, and achieves SOTA on OVO-Bench, ESTP-Bench, EgoSchema, and QAEgo4D.

Original post →

More from Research

Research channel →