Reproduce “Context Compaction for World Models” 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

Hayder Tirmazi

The authors introduce a harness that compacts session histories into walks through previously visited rooms, extending context without changing model weights. On a generated room-navigation benchmark, compaction raises context-retrieval gain to 0.39–0.51 from 0.03–0.06 for revisits 128–511 steps back. The results concern small models and do not establish why sliding-window lower blocks sharply reduce retrieval during dreaming.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.