Reproduce “ACLArena: Agent Continue Learning in Multi-stage Post-training” 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
Haixin Wang, Xiaoxuan Wang, Junkai Zhang, Han Zhang, Renliang Sun, Alexander K Taylor, Yidan Shi, Haoran Deng and 4 more
MLE combines high-quality trajectories from task specialists into a shared model, then trains separate low-rank adapters with reinforcement-learning for each task. In a fixed Math-to-Search-to-E-commerce-to-Instruction-Following curriculum on Qwen3-8B-Base, it raises out-of-domain multi-hop Search average to 38.6±0.1, from 25.0±1.5 for Seq-Final, while IF-Eval is 85.0±0.3 versus 84.8±0.1; evidence is limited to this model and curriculum, and does not isolate the cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.