Reproduce “Memory Internalization via Test-Time Training for Long-Video Understanding” 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
Haonan Zhao, Jiaxuan Chen, Qingyang Liu, Shuochen Chang, Siyu Sun, Jiangtong Li, Li Niu
MIT³ (Memory Internalization via Test-Time Training) prepares a video-specific LoRA memory so later questions can be answered from text alone. It selects layers for writing the memory and aligns a text-only student’s representation changes with those of a video-conditioned teacher. On LVBench videos longer than one hour, Qwen3-VL-32B reaches 50.1% accuracy, above matched NKR-style and direct-inference baselines in this setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.