Reproduce “dKV Forcing: Separating Generation and Memory Representations through a Differentiable KV Cache” 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
Zhiqi Li, Bo Zhu
dKV Forcing trains causal video generators to write cache representations for future use: later segments’ losses update earlier writes, with a writer-specific low-rank residual. Immutable pages and local rematerialization enable this gradient in one traversal. On a 1.3B model trained on five-second clips, the residual scored 59.8 versus SGF’s 51.3 on 240-second VBench-Long aesthetic quality, limited to this setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.