Reproduce “NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness” 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

NeoHorse Team, Guoliang Cao, Guohao Dai, Tianyu Guo, Kai Han, Hailin Hu, Zihan Jiang, Xiang Kuang and 29 more

NeoHorse-1 turns execution records from tool-using agents into post-training examples and feedback for future training. It uses router estimates of capability demand to organize examples and adjust the training mixture. Across ten agentic, coding, and instruction-following benchmarks, its 4B and 9B models raise unweighted average scores from 58.94 to 64.87 and 65.60 to 69.04, respectively; gains across generations remain untested.

Summary by alphaXiv.

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