A new paper submitted to the EgoLongQA track of the Wearable-AI Challenge at ECCV 2026 describes a system that took first place in the ≤2B parameter division with a score of 0.8279 on the held-out test set. The entry, titled Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model, is a single 2B vision-language model that answers multiple-choice questions about ten-minute egocentric videos in one greedy forward pass.
Rather than distilling an entire agentic pipeline, the authors distilled only the junior perception module of a tool-using agent into a small student model. The teacher traces used for training were filtered to those that answered correctly. According to the paper, the resulting student reaches 89% of the accuracy of the large agentic pipeline while using just 1.1% of its parameters, lifting a base model from 27.1% to 81.4% on the authors' held-out questions.
Meeting the divisional parameter limit required a technical workaround. The 2B backbone actually contains 2.2132B parameters, above the ≤2B threshold. To make the entry admissible, the team pruned the multilingual embedding table from 248,320 rows down to 143,469 rows, arriving at 1.9985B parameters with provably identical logits on the retained rows.
The work sits alongside a growing body of 2026 research on long-video understanding and egocentric assistance, including related papers on proactive egocentric assistance, temporal self-distillation, and on-policy self-distillation for long-video understanding.