CuSearch: Curriculum Rollout Sampling for Agentic RAG Training

CuSearch is a proposed framework for implementing a curriculum rollout sampling aimed at training agentic retrieval-augmented generation (RAG) systems through Reinforcement Learning with Verifiable Rewards (RLVR). Unlike current approaches that regard all trajectories as equal, deeper-search trajectories offer more retrieval decision points and richer supervision. To tackle the increasing heterogeneity in search depth during training, CuSearch employs Search-Depth Greedy Allocation (SDGA) to redistribute the update budget towards these deeper-search trajectories. This framework is detailed in the arXiv paper numbered 2605.11611.