Research map / Graph learning / Knowledge and temporal graphs
ReMindRAG: Low-Cost LLM-Guided Knowledge Graph Traversal for Efficient RAG
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 ReMindRAG, a training-free KG-RAG system where LLM-guided graph traversal is memorized as closed-form edge-embedding updates, so repeated or similar queries need fewer LLM calls.
Problem
KG-RAG systems face a trade-off: graph-search methods are cheap but less accurate, while LLM-guided graph traversal is accurate but costs many LLM prompt tokens and inference time.
Approach
Store LLM traversal experience as edge embeddings in the KG, updated in closed form (enhance edges on effective paths along the query embedding, penalize edges on ineffective paths via a norm-dependent weight function). At query time, these embeddings guide DFS-based subgraph expansion (memory replay) before any LLM-guided traversal, so the LLM is called only when needed.
Key results
- ReMindRAG reaches 57.04% (GPT-4o-mini) on LooGLE long-dependency QA versus 39.60% for HippoRAG2 and 27.78% for Plan-on-Graph.
- With multi-turn memorization, token consumption on LooGLE long-dependency QA drops from 14.91K (no memorization) to 6.71K after 3 turns, and the paper reports an average 58.8% token reduction after multi-turn memorization.
- On Multi-Hop QA with GPT-4o-mini, ReMindRAG achieves 74.22% accuracy versus 68.04% for HippoRAG2 with no memorization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ReMindRAG instead of the strongest reported baseline | KG-RAG · LooGLE Long Dependency QA | accuracy: 39.60% -> 57.04%; 53.02% -> 59.73% | Helped |
| Use ReMindRAG instead of the strongest reported baseline | KG-RAG · HotpotQA Multi-Hop QA | accuracy: 68.04% -> 74.22%; 64.95% -> 79.38% | Helped |
| Use ReMindRAG instead of the strongest reported baseline | KG-RAG · LooGLE Simple QA | accuracy: 73.08% -> 76.67%; 71.28% -> 77.01% | Helped |
| Add the Contextual Skeleton Network (ablation) | ReMindRAG · LooGLE Long Dependency QA | accuracy: 51.01% -> 57.04% | Helped |
| Add the Contextual Skeleton Network (ablation) | ReMindRAG · HotpotQA Multi-Hop QA | accuracy: 74.22% -> 74.22% | No clear effect |
| Add the Contextual Skeleton Network (ablation) | ReMindRAG · LooGLE Simple QA | accuracy: 73.07% -> 76.67% | Helped |
| Add memory replay and increase memorization from none to 3 turns (ablation) | ReMindRAG · LooGLE Long Dependency QA, same query | tokens per query: 14.91K -> 6.71K | Helped |
| Add memory replay and increase memorization from none to 3 turns (ablation) | ReMindRAG · HotpotQA Multi-Hop QA, same query | accuracy: 74.22% -> 87.62% | Helped |
| Add memory replay and increase memorization from none to 3 turns (ablation) | ReMindRAG · HotpotQA Multi-Hop QA, same query | tokens per query: 10.16K -> 5.89K | Helped |
| Increase memorization from 1 turn to 3 turns (ablation) | ReMindRAG · HotpotQA Multi-Hop QA, similar query | accuracy: 75.53% -> 78.72% | Helped |
| Increase memorization from 1 turn to 3 turns (ablation) | ReMindRAG · HotpotQA Multi-Hop QA, different query | accuracy: 72.34% -> 77.66% | Helped |
| Increase memorization from 1 turn to 3 turns (ablation) | ReMindRAG · LooGLE Long Dependency QA, similar query | accuracy: 55.03% -> 57.25% | Mixed |
Surprising findings
- Accuracy on the Multi-Hop and Long Dependency tasks improves after repeated memorization, with the paper attributing this to self-correction of traversal paths without access to ground-truth answers.
- Truncating embedding dimensions down to 12 causes only modest accuracy drops, which the authors attribute to Matryoshka-style embeddings.
Limitations
- The theoretical capacity analysis assumes embedding dimension approaches infinity, with the practical relevance argued only heuristically for dimensions around 100.
- The LooGLE questions were reformatted into multiple-choice form with GPT-4o, and evaluation relies on GPT-4o as judge, which may bias results.
- Only three benchmark settings and two LLM backbones are evaluated, and the Different Query setting is only available for Multi-Hop QA.
Experimental setup
Datasets: LooGLE HotpotQA
Benchmarks: LooGLE HotpotQA
Metrics: answer accuracy average tokens per query
Baselines: BM25 NaiveRAG GraphRAG LightRAG HippoRAG2 Plan-on-Graph MemoRAG
Builds on: GraphRAG LightRAG HippoRAG2 Plan-on-Graph Matryoshka Representation Learning
Closest papers by meaning
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.89)
- Can Knowledge-Graph-based Retrieval Augmented Generation Really Retrieve What You Need? (NeurIPS 2025 · similarity 0.88)
- Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
- GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.88)
- StructRAG: Boosting Knowledge Intensive Reasoning of LLMs via Inference-time Hybrid Information Structurization (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.