AgenticRag-R1: Agentic Reinforcement Learning with Stack Memory for Multi-Step Reasoning, Retrieval and Memorizing
The paper proposes a RAG agent that uses a stack-like memory system to manage longer reasoning paths.
AgenticRag-R1 combines reinforcement learning with fine-grained actions for reasoning, retrieval, and memory updates. The authors say its reward design targets individual action choices rather than only whole trajectories, aiming to reduce weak credit assignment. They report stronger results than baselines across multi-hop, open-domain, and agentic reasoning benchmarks using multiple backbone sizes. Code is described as anonymously available. ArXiv · AI/CL/LG's note
AgenticRag-R1 combines reinforcement learning with fine-grained actions for reasoning, retrieval, and memory updates. The authors say its reward design targets individual action choices rather than only whole trajectories, aiming to reduce weak credit assignment. They report stronger results than baselines across multi-hop, open-domain, and agentic reasoning benchmarks using multiple backbone sizes. Code is described as anonymously available. ArXiv · AI/CL/LG's note
score 5