CuSearch framework enhances agentic RAG training with curriculum sampling

Researchers have developed CuSearch, a new framework for training agentic retrieval-augmented generation (RAG) systems using Reinforcement Learning with Verifiable Rewards (RLVR). This method addresses the issue of uniformly sampling trajectories by prioritizing deeper-search trajectories, which offer more informative supervision for the retrieval sub-policy. CuSearch utilizes Search-Depth Greedy Allocation (SDGA) to dynamically allocate update budgets towards these deeper trajectories, leading to improved performance, with experiments showing up to an 11.8 exact-match point improvement over standard GRPO on the ZeroSearch benchmark. AI

IMPACT This research could lead to more efficient training of agentic RAG systems, improving their ability to retrieve and utilize information.

RANK_REASON The cluster contains a research paper detailing a new method for training AI systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• agentic retrieval-augmented generation
• CuSearch
• GRPO
• Jianghan Shen
• Reinforcement Learning with Verifiable Rewards
• RLVR
• Search-Depth Greedy Allocation
• ZeroSearch

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →