EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos
EgoTools tests whether video models can reason through tool use from a first-person view, and the answer is still uneven.
The paper introduces EgoTools-Data: 100 hours of egocentric, tool-centered recordings with audio, captions, narrations, and 3D information. It also adds EgoTools-Bench, a 1,000-question benchmark across perception, geometry, procedure, and causal reasoning. Gemini-3.1-Pro reaches 66.9% overall accuracy, but drops to 51.7% on Perception & Grounding. Fine-tuning Qwen3-VL-8B-Instruct on the dataset improves benchmark accuracy from 50.0% to 60.9% under source-video separation. HF Daily Papers' note
The paper introduces EgoTools-Data: 100 hours of egocentric, tool-centered recordings with audio, captions, narrations, and 3D information. It also adds EgoTools-Bench, a 1,000-question benchmark across perception, geometry, procedure, and causal reasoning. Gemini-3.1-Pro reaches 66.9% overall accuracy, but drops to 51.7% on Perception & Grounding. Fine-tuning Qwen3-VL-8B-Instruct on the dataset improves benchmark accuracy from 50.0% to 60.9% under source-video separation. HF Daily Papers' note
score 4