Reproduce “Latent Compaction with Test-Time Training” 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
Yingfa Chen, Xingyu Shen, Chaojun Xiao, Xu Han
Latent Compaction learns a document-specific compact key–value cache before questions arrive, using generic prompts while keeping model weights frozen. It fits attention outputs with selected preservers fixed, then refines the cache against full-cache predictions. On Phonebook Lookup, Qwen3-1.7B at 16× compression averages 80.67% accuracy, versus 33.42% when preservers train. This is a tested lookup result; efficiency evidence concerns logical KV-pair budgets, not measured physical memory or inference speed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.