Research map / Large language models / Retrieval and knowledge

NeuroPath: Neurobiology-Inspired Path Tracking and Reflection for Semantically Coherent Retrieval

NeurIPS 2025 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NeuroPath, a graph-based RAG framework that uses an LLM to track and prune semantic paths over a knowledge graph (inspired by hippocampal place-cell preplay/replay) and then runs a second-stage retrieval with the intermediate reasoning, improving multi-hop QA retrieval.

Problem

Graph-based RAG for multi-hop QA loses semantic coherence and introduces irrelevant noise during node matching and subgraph construction, while iterative RAG lacks explicit modeling of associations between knowledge.

Approach

Model KG entities as place cells and triples as place fields; an LLM goal-directedly expands and prunes candidate semantic paths from seed nodes using expansion requirements from the previous hop, then a second-stage retrieval combines the query with the last-hop reasoning chain and expansion requirements to complete missing information.

Key results

• NeuroPath (Zero-Shot, GPT-4o-mini + Contriever) reaches average R@2 66.9 and R@5 81.9, versus HippoRAG 2 at 56.5 and 71.0 on the same setup.
• Compared with iter-based methods, NeuroPath reduces token consumption by 22.8% on average during retrieval while improving recall (average gains of 8.6% R@2 and 10.2% R@5 claimed).
• On MuSiQue QA with Contriever, NeuroPath achieves EM 31.4 / F1 44.3, the best among compared methods; it trails HippoRAG 2 slightly on HotpotQA QA.

Measured effects reported in the paper

Surprising findings

• HippoRAG 2 outperforms NeuroPath on HotpotQA QA metrics despite lower retrieval recall, which the authors attribute to answer guessing shortcuts.
• LightRAG and PathRAG perform worse than even naive RAG on multi-hop factual QA.

Limitations

• Heavy reliance on LLM calls for path filtering and expansion creates a time bottleneck (retrieval 6.7 s per question vs 4.7-5.6 s for iter-based methods).
• Low granularity of triple segments may split simple questions into unnecessary multi-hop steps.
• Does not optimize the corpus (e.g., via summarization), so it may struggle with sense-making tasks requiring abstraction.

Experimental setup

Datasets: MuSiQue 2WikiMultiHopQA HotpotQA PopQA Natural Questions MultiHop-RAG

Benchmarks: MuSiQue 2WikiMultiHopQA HotpotQA PopQA Natural Questions MultiHop-RAG

Metrics: recall@2 recall@5 exact match F1 token consumption

Baselines: BM25 Contriever BGE-M3 IRCoT Iter-RetGen HippoRAG HippoRAG 2 LightRAG PathRAG

Builds on: HippoRAG LightRAG PathRAG IRCoT Iter-RetGen

Closest papers by meaning

• HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (NeurIPS 2024 · similarity 0.88)
• Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.88)
• GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.88)
• RAGRouter: Learning to Route Queries to Multiple Retrieval-Augmented Language Models (NeurIPS 2025 · similarity 0.87)
• Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.87)
• From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025 · similarity 0.87)