Small vision language model answers questions about long egocentric videos
Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model
Computer Vision and Pattern Recognition
Summary
Watching and understanding long videos from a first-person view is hard for computers. The paper describes a small model made by teaching it from a much bigger system, letting it answer questions about ten-minute videos. This small model is very efficient, using just a tiny fraction of the bigger model’s size, but still performs well. The authors also shrink the model further by removing unused language parts without hurting its ability to answer questions.
What this means in practice
- •For wearable device developers: Build compact AI systems that answer questions about long first-person videos efficiently on limited hardware.$Commercial implications: Enables practical AI-powered wearable devices that analyze user video streams with low resource demands, improving real-time interaction.
- •For mobile app developers: Implement lightweight video understanding features answering multiple-choice questions without large computational resources.
Authors
Logesh Kumar Umapathi
Abstract
We describe our entry to the EgoLongQA track of the Wearable-AI Challenge in ECCV 2026, which placed first in the <=2B parameter division with 0.8279 on the held-out test set. Our system is a single 2B vision-language model that answers multiple-choice questions about ten-minute egocentric videos in one greedy forward pass; It is obtained by distilling the junior perception module of a tool-using agentic pipeline, not the agent itself into a small student, using teacher traces filtered to those that answered correctly. it reaches 89% of the accuracy of the large agentic pipeline using 1.1% of its parameters. This raises a 27.1% base model to 81.4% on our held-out questions. The 2B backbone has 2.2132B parameters and therefore over the divisional limit, to make the entry admissable we prune the multilingual embedding table from 248,320 to 143,469 rows, reaching 1.9985B with provably identical logits on retained rows.