Reproduce “Language Model Harnesses Can Be Compositional Generalizers” 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
Alex L. Zhang, Souradip Chakraborty, Omar Khattab
The paper trains a recursive language model harness to handle long tasks by keeping context outside the prompt, using code to inspect it, and delegating smaller queries to model calls. On eight held-out long-context benchmarks, RLM-27B averages 58.5 versus 46.1 for Qwen3.8-27B in the same harness; this is a benchmark-score gain, not a compute or time result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.