Reproduce “Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Logesh Kumar Umapathi
Ambient distills the perception component of a tool-using egocentric video QA system into a single-pass vision-language model with 1.9985B parameters. Trained on 2,936 oracle-filtered teacher traces, the compact model reaches 81.4% on a 70-video held-out set, 56.6% on a 500-question out-of-distribution evaluation with balanced synthetic data, and 0.8279 in the EgoLongQA ≤2B challenge division.
Key insights
- A junior perception module can be a more compatible distillation target than the full agent, because its fixed input-output interface matches single-pass student inference even though the senior tool-using component does not.
- High-quality oracle-filtered real traces are much more valuable than synthetic questions for in-distribution performance, whereas balanced synthetic data improves out-of-distribution accuracy and reduces answer-option bias.
- More frames are not automatically beneficial for small multimodal models: increasing the budget from 100 to 400 frames lowered performance in one ≤2B comparison, suggesting a capacity bottleneck.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.