Phi-agent: A Dynamic In-Graph Reasoning Agent Inspired by the Prefrontal–hippocampal Interaction

ICLR 2026 Conference Submission16572 Authors

Keywords: Large Language Models, In-context Learning, Retrieval-Augmented Generation, Neuro-inspired AI, Reasoning Agent

Abstract: Graph-based Retrieval-Augmented Generation has demonstrated strong performance in multi-hop reasoning and cross-document evidence integration. However, existing methods typically rely on static, one-shot retrieval strategies, lacking the ability to assess evidence sufficiency or dynamically construct context—thereby limiting their effectiveness in complex reasoning tasks. To address these limitations, we propose Phi-agent, a brain-inspired iterative in-graph retrieval agent. Motivated by the hippocampus–prefrontal interaction in cognitive neuroscience, Phi-agent operates in a "retrieve–reason–re-retrieve" loop, enabling proactive refinement of contextual evidence through iterative questioning and reasoning. We further introduce a joint reward mechanism that simultaneously optimizes both reasoning quality and retrieval trajectory. To support reinforcement learning, we curate a high-quality dataset of 7,405 annotated samples and post-train Qwen3-1.7B using the Group Relative Policy Optimization (GRPO) algorithm. Experiments on HotpotQA, MuSiQue, and 2WikiQA show that Phi-agent significantly outperforms existing GraphRAG baselines, achieving state-of-the-art performance. Ablation studies confirm the essential role of the iterative in-graph retrieval loop and joint reward design in enabling these improvements.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 16572