UCS-Bench

ICML 2026 Keep It in Mind: User Centric Continual Spatial Intelligence Reasoning in Egocentric Video Streams

UCS-Bench is a benchmark for evaluating user-centric continual spatial intelligence in streaming egocentric videos.

The dataset contains 170+ hours of egocentric visual observations and 8.1K+ timestamped questions. It is designed to test whether models can understand dynamic spatial environments from a user's point of view, maintain long-term spatial memory, and answer questions grounded in the user's real-time movement and location.

Unlike standard video QA benchmarks, UCS-Bench focuses on long-horizon spatial reasoning in continuous egocentric streams, including:

• tracking object locations over time;
• recalling previous observations and interactions;
• resolving viewpoint-induced spatial ambiguity;
• reasoning about dynamic scenes from the user's perspective;
• aligning spatial memory with the user's changing location.

Links

• Code: https://github.com/cocowy1/UCS-Bench
• Paper: [ICML 2026] https://arxiv.org/abs/2606.15200

Dataset Format

Each video is paired with metadata in metadata.jsonl.

Example:

{"file_name": "video001.mp4",  "label": "xxx"}