Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model

Hugging Face Daily Papers Papers
vision-language-model long-video-understanding model-distillation egocentric-video parameter-pruning eccv-2026 competition-winner

Summary

A 2B vision-language model distilled from a tool-using agent achieves first-place performance on long egocentric video question answering by pruning its multilingual embedding table to meet parameter limits, reaching 89% accuracy of the larger pipeline with only 1.1% of parameters.

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.
Original Article
View Cached Full Text

Cached at: 09/14/26, 06:36 PM

Paper page - Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model

Source: https://huggingface.co/papers/2609.07154

Abstract

A 2B vision-language model distilled from a tool-using agent achieves first-place performance on long egocentric video question answering by pruning its multilingual embedding table to meet parameter limits.

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 2Bvision-language modelthat answersmultiple-choice questionsabout ten-minuteegocentric videos in one greedy forward pass; It is obtained by distilling the junior perception module of atool-using agentic pipeline, not the agent itself into a small student, usingteacher tracesfiltered 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 themultilingual embedding tablefrom 248,320 to 143,469 rows, reaching 1.9985B with provably identical logits on retained rows.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2609\.07154

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.07154 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.07154 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.07154 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos

Hugging Face Daily Papers

EgoTools introduces a 100-hour egocentric tool-use corpus and a 1,000-question diagnostic benchmark, showing current multimodal models (e.g., Gemini-3.1-Pro at 51.7% on Perception & Grounding) still struggle with tool-centric embodied reasoning, while fine-tuning Qwen3-VL-8B on the data lifts accuracy from 50.0% to 60.9%.

MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories

Hugging Face Daily Papers

MemLife introduces a multimodal memory system for long-term egocentric videos that builds entity-grounded text episodes and retrieves them with a time-indexed agentic reader, improving over training-free baselines by 4.6-12.0% on long-horizon benchmarks. A reinforcement learning framework called MemOpt further optimizes the memory writer for faithfulness and retrievability, yielding consistent 2.7-5.0% gains.

SuperMemory-VQA: An Egocentric Visual Question-Answering Benchmark for Long-Horizon Memory

Hugging Face Daily Papers

SuperMemory-VQA is a new egocentric VQA benchmark featuring 52.9 hours of AI-glasses footage and 4,853 QA pairs designed to evaluate AI assistants on long-horizon memory tasks spanning object recall, intent, timelines, and conversations. Benchmarking reveals existing agentic frameworks and LLMs remain far from reliable on these real-world memory challenges.

Motif-Video 2B: Technical Report

Hugging Face Daily Papers

Motif-Video 2B is a 2B parameter text-to-video generation model that achieves 83.76% on VBench, surpassing Wan2.1 14B while using 7x fewer parameters and trained on fewer than 10M clips with less than 100,000 H200 GPU hours. The model uses a specialized architecture with shared cross-attention and a three-part backbone to separate prompt alignment, temporal consistency, and detail refinement.