Reproduce “Don't Lose the Thread: Empowering Long-Horizon LLM Agents with Cognitive Resource Self-Allocation” 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
Zheng Zhu, Xiaozhuan Liang, Xi Chen, Bei Yu, Jiaya Jia
CORAL lets web agents save plans and verified facts, then reset cluttered reasoning histories into new episodes using that memory. It trains agents to manage these checkpoints through supervised fine-tuning and multi-episode reinforcement learning. On GAIA, Qwen3-8B CORAL reaches 42.8% average accuracy while using 87.4% fewer input tokens per trajectory than ReAct; its average is below CK-Pro-8B’s 43.7%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.