Reproduce “Tools Need Not Speak in Text: Latent Context Compression for Agents” 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

Yuetong Liu, Zhehao Zhou, Yitong Zhang, Haorui Chen, Jia Li

Coca-Cola replaces web-search observations with learned latent vectors trained to preserve an agent’s next-action behavior, rather than reconstruct the source text. Across four web-search QA benchmarks and two model configurations, it reduces total token use by 70.01%–89.24% and runtime by 24.45%–55.97%, while improving answer scores over uncompressed agents; results are limited to these settings.

Summary by alphaXiv.

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