New PAR²-RAG framework boosts multi-hop question answering accuracy by up to 23.5% through planned retrieval and adaptive reasoning

PAR²-RAG addresses brittleness in LLMs by separating evidence coverage from commitment in a two-stage retrieval and reasoning process

System uses breadth-first anchoring to build high-recall evidence frontier, followed by depth-first refinement with sufficiency controls

Outperforms existing baselines across four MHQA benchmarks, achieving up to 23.5% higher accuracy compared to IRCoT

Solves problems of iterative retrieval systems locking onto low-recall trajectories and static planning approaches that fail to adapt to changing evidence