AgenticRAG-R1: RL Framework with Stack Memory for Reasoning Agents

_reachsumit · x · 2026-09-01

The paper proposes AgenticRAG-R1, an RL framework for training RAG agents.

Innovation: Introduces a "stack memory" and fine-grained action space (plan, search, backtrack) to deeply integrate reasoning, retrieval, and memory.
Solution: Addresses weak reward assignment and short-horizon bias in existing RL methods that rely on coarse-grained actions and trajectory-level rewards.
Performance: Outperforms baselines on multi-hop, open-domain, and agentic reasoning benchmarks, learning more robust and interpretable behaviors.

More from coding & agent