MM-Lifelong: 181 Hours of Video for Long-Term Multimodal Reasoning

MM-Lifelong provides 181.1 hours of video footage across three domains for multimodal lifelong understanding, released by CG-Bench in 2026. The dataset includes 1,289 questions and 1,810 clue intervals designed to test reasoning over extended temporal spans. It specifically targets long-context video comprehension with questions requiring up to 10+ hours of dependency reasoning.

Use Cases

• Visual Question Answering (VQA) using the 1,289 questions to test long-form video understanding
• Temporal grounding by locating the 1,810 clue intervals within the footage
• Long-context reasoning evaluation for dependencies spanning 1-10+ hours of video

Strengths

• 181.1 hours of video footage across three domains
• 127 questions requiring reasoning over ultra-long dependencies exceeding 10 hours
• 1,810 distinct clue intervals provided for temporal grounding

Limitations

• Small sample size of 1,289 questions relative to total video duration
• Limited domain diversity covering only three specific areas
• High computational requirements for processing 10+ hour video contexts

Provenance

Source

CG-Bench (Arxiv: 2603.05484)

Freshness

Last updated March 2026

Geography

United States