Reproduce “Mosaic: Multi-Session Agentic Reinforcement Learning” 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
Yilong Xu, Minrui Xu, Xinyu Zhou, Xiang Long, Haochen Tan, Zhijiang Guo
Mosaic trains web-research agents to handle tasks split across context-compaction or subagent sessions by combining session-count-aware trajectory advantages with session-level changes in the policy’s likelihood of the correct answer. On BrowseComp, Qwen3.5-4B reaches 38.86% accuracy with 2.35 sessions per trajectory under compaction, versus broadcasting’s 35.31% and 3.17 sessions; evidence is limited to this model and benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.