Beyond Similarity through Zero-Token Geometric Graphs for Multi-Hop RAG
G3RAG uses geometric gain graphs for zero-token construction, enhancing multi-hop RAG with a 4.26 F1 score increase.
Key Findings
Methodology
G3RAG employs geometric gain graphs, defining edge weights using directional consistency and orthogonality of document representations. A density-aware topological penalty suppresses highly connected hubs, while single-step controlled diffusion expands from filtered query seeds to complementary evidence.
Key Results
- G3RAG achieved a 5.76-point F1 score increase on the MusiQue dataset, with an average performance boost of 4.26 points, demonstrating its superiority in multi-hop RAG.
- On 2WikiMultiHopQA and HotpotQA datasets, G3RAG outperformed all evaluated graph-based baselines in both Nv-embed-v2 and Qwen3-8B-embed settings.
- G3RAG eliminated the token cost of graph construction incurred by entity-based methods, significantly enhancing efficiency.
Significance
G3RAG supports efficient multi-hop evidence discovery through geometric structures without LLM-based graph construction, addressing high costs and noise propagation issues in traditional methods. Its innovative graph construction approach provides new insights for multi-hop RAG, with significant academic and industrial implications.
Technical Contribution
G3RAG achieves zero-token construction via geometric gain graphs, offering new theoretical guarantees and engineering possibilities compared to existing SOTA methods. Its geometric gain scoring mechanism and topological penalty strategy enhance evidence retrieval efficiency in multi-hop RAG.
Novelty
G3RAG is the first to introduce geometric gain graphs in multi-hop RAG, avoiding the costs of LLM calls and token generation. Compared to entity-based methods, it significantly improves construction efficiency and performance.
Limitations
- In handling extremely complex queries, G3RAG may still be affected by the similarity trap, leading to connections to locally homogenized redundant documents.
- Further optimization of topological penalty parameters is needed to adapt to different dataset characteristics.
Future Work
Future work could explore G3RAG's applications in other tasks, such as information retrieval and QA systems, and optimize its parameter settings across different datasets to further enhance performance.
AI Executive Summary
Multi-hop retrieval-augmented generation (RAG) requires evidence that remains relevant to a query while introducing enough novelty to bridge semantic gaps. Traditional dense retrieval methods tend to focus on semantically similar documents, while graph-based alternatives rely on costly large language model (LLM) entity extraction and may propagate through noisy connections. We introduce Geometric Gain Graph RAG (G3RAG), a document-only framework whose offline graph construction uses no LLM calls or generated tokens. G3RAG assigns each edge a geometric gain score, capturing directional consistency and orthogonality between document representations. A density-aware topological penalty suppresses highly connected hubs, while single-step controlled diffusion expands from filtered query seeds toward complementary evidence. We evaluate G3RAG on MusiQue, 2WikiMultiHopQA, and HotpotQA using Nv-embed-v2 and Qwen3-8B-embed. G3RAG achieves the best average F1 and answer-document hit rate among evaluated graph-based baselines in both embedding settings, with gains of up to 4.26 F1 points in average performance and 5.76 points on MusiQue. It also removes the graph-construction token cost incurred by entity-based methods. These results show that geometric structure can support efficient multi-hop evidence discovery without LLM-based graph construction. Code is available at https://anonymous.4open.science/r/G3RAG-99D9/.
Deep Analysis
Background
Multi-hop retrieval-augmented generation (RAG) is a core paradigm for modern generative AI, aiming to mitigate LLM hallucinations through external knowledge. However, traditional dense retrieval methods often fall into the 'similarity trap' when handling multi-hop query scenarios, relying heavily on surface-level semantic matching, leading to highly homogeneous document retrieval that hinders the acquisition of novel, critical supporting evidence.
Core Problem
The core problem faced by multi-hop RAG is how to introduce enough novelty while maintaining query relevance to bridge semantic gaps. Traditional methods often rely on costly LLM entity extraction and may propagate through noisy connections, resulting in high costs and limited performance improvements.
Innovation
G3RAG's core innovation lies in introducing geometric gain graphs, defining edge weights using directional consistency and orthogonality of document representations, achieving zero-token construction. This approach avoids the costs of LLM calls and token generation while suppressing highly connected hubs through a density-aware topological penalty.
Methodology
- �� G3RAG constructs geometric gain graphs, defining edge weights using directional consistency and orthogonality.
- �� A density-aware topological penalty suppresses highly connected hubs.
- �� Single-step controlled diffusion expands from filtered query seeds to complementary evidence.
Experiments
We evaluated G3RAG on MusiQue, 2WikiMultiHopQA, and HotpotQA using Nv-embed-v2 and Qwen3-8B-embed. The experimental design includes comparisons with baseline methods, performance evaluations across different embedding models, and ablation studies.
Results
G3RAG achieved a 5.76-point F1 score increase on the MusiQue dataset, with an average performance boost of 4.26 points, demonstrating its superiority in multi-hop RAG. G3RAG outperformed all evaluated graph-based baselines across all datasets.
Applications
G3RAG can be applied in multi-hop QA systems, information retrieval, and other scenarios where efficient evidence discovery is crucial, offering significant advantages in complex tasks.
Limitations & Outlook
G3RAG may still be affected by the similarity trap when handling extremely complex queries. Future work could explore its applications in other tasks and optimize its parameter settings across different datasets.
Plain Language Accessible to non-experts
Imagine you're in a library searching for a book. Traditional methods are like only looking at book titles to find books, resulting in finding very similar books. G3RAG's method is like finding books based on their content and themes, allowing you to find more diverse yet relevant books. This way, G3RAG can find more useful information without additional costs.
ELI14 Explained like you're 14
Imagine you're playing a treasure hunt game. Traditional methods are like only following map markers to find treasures, resulting in finding very similar treasures. G3RAG's method is like finding treasures based on map terrain and clues, allowing you to find more diverse yet interesting treasures. This way, G3RAG can find more exciting treasures without spending more time.
Glossary
Geometric Gain
A method defining edge weights using directional consistency and orthogonality of document representations.
Used in G3RAG's graph construction to balance relevance and novelty.
Zero-Token
A graph construction method that does not require LLM calls or token generation.
G3RAG achieves zero-token construction through geometric gain.
Density-Aware Topological Penalty
A strategy to suppress highly connected hubs to maintain graph diversity.
Used in G3RAG's graph construction to avoid the similarity trap.
Controlled Diffusion
A diffusion method expanding from filtered query seeds to complementary evidence.
Used in G3RAG's evidence retrieval to enhance efficiency.
Multi-Hop RAG
An AI method enhancing generation through multi-hop retrieval.
G3RAG aims to improve multi-hop RAG performance.
Open Questions Unanswered questions from this research
- 1 How to optimize G3RAG's topological penalty parameters across different datasets to enhance performance.
- 2 Explore G3RAG's applications in other tasks, such as information retrieval and QA systems.
Applications
Immediate Applications
Multi-Hop QA Systems
G3RAG can enhance multi-hop QA system performance, especially in complex tasks requiring efficient evidence discovery.
Long-term Vision
Information Retrieval
G3RAG can be applied in information retrieval, offering significant advantages in complex tasks requiring efficient evidence discovery.
Abstract
Multi-hop retrieval-augmented generation (RAG) requires evidence that remains relevant to a query while introducing enough novelty to bridge semantic gaps. Dense retrieval tends to concentrate on semantically similar documents, whereas graph-based alternatives often depend on costly Large Language Model (LLM) entity extraction and may propagate through noisy connections. We introduce Geometric Gain Graph RAG (G$^3$RAG), a document-only framework whose offline graph construction uses no LLM calls or generated tokens. G$^3$RAG assigns each edge a geometric gain score, $\cosθ\cdot \sinθ$, that jointly captures directional consistency and orthogonality between document representations. A density-aware topological penalty suppresses highly connected hubs, while single-step controlled diffusion expands from filtered query seeds toward complementary evidence. We evaluate G$^3$RAG on MusiQue, 2WikiMultiHopQA, and HotpotQA using Nv-embed-v2 and Qwen3-8B-embed. G$^3$RAG obtains the best average F1 and answer-document hit rate among the evaluated graph-based baselines in both embedding settings, with gains of up to 4.26 F1 points in average performance and 5.76 points on MusiQue. It also removes the graph-construction token cost incurred by entity-based graph methods. These results show that geometric structure can support efficient multi-hop evidence discovery without LLM-based graph construction. Code is available at https://anonymous.4open.science/r/G3RAG-99D9/