Reproduce “Language Models Learn to Improve Language 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
Minhua Lin, Zijun Wang, Zewen Liu, Zhaotian Weng, Bing He, Zhan Shi, Benoit Dumoulin, Xiang Zhang and 2 more
LM2LM trains a language-model proposer using rewards from a frozen, record-grounded research world model, then selectively executes highly rated proposals and adds measured outcomes to refresh its context. With the Qwen3 family withheld from training and grounding, Qwen3.8-27B scored 53.72 Avg@4 versus 21.63 for GPT-6-sol using 300 out-of-family records. This shows transfer in the tested proposal-level setting, not arbitrary research environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.