AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Geometric Gain Graph: Zero-Token Graph Construction for Multi-Hop RAG

Zeliang Li, Xiaofen Xing, Wenyu Tao, Kailing Guo, Xiangmin Xu

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 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 Geometric Gain Graph, a zero-token graph construction method for multi-hop RAG that balances relevance and novelty without LLM-based entity extraction.

Problem

Multi-hop reasoning in RAG fails because dense retrievers collapse onto similar surface-level documents (the "Similarity Trap"), while Graph RAG methods explore novel documents via heuristic entity extraction at prohibitive LLM cost.

Approach

Builds the retrieval graph geometrically from embeddings so that relevance and novelty are balanced without any LLM calls for entity extraction, giving zero-token graph construction.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add topological degree penalty to graph edges (ablation)G3RAG · MuSiQue (Nv-embed-v2)F1: 50.65 -> 52.28Improved
add topological degree penalty to graph edges (ablation)G3RAG · HotpotQA (Nv-embed-v2)F1: 77.60 -> 78.00Improved
use single-step controlled diffusion instead of Personalized PageRank (ablation)G3RAG · Average of 3 datasets (Nv-embed-v2)F1: 65.65 -> 66.49Improved
add LLM filtering of seed nodes (ablation)G3RAG · Average of 3 datasets (Nv-embed-v2)F1: 60.53 -> 66.49Improved
add geometric graph diffusion on top of LLM-filtered naive retrieval (ablation)Naive RAG with Filter_llm · Average of 3 datasets (Nv-embed-v2)F1: 63.39 -> 66.49Improved
use single-step diffusion instead of greedy walk or beam search (ablation)G3RAG · Average of 3 datasets (Nv-embed-v2)F1: 64.79 (greedy), 64.87 (beam) -> 66.49Improved
use global density estimate instead of local 10-NN density for penalty (ablation)G3RAG · Average of 3 datasets (Nv-embed-v2)F1: 64.97 -> 66.49Improved
use geometric gain graph instead of MMR or DPP diversificationNaive RAG · Average of 3 datasets (Nv-embed-v2)F1: MMR 52.47, DPP 60.93 -> 66.49Improved
add MMR diversification to naive dense retrievalNaive RAG · Average of 3 datasets (Nv-embed-v2)F1: 60.62 -> 52.47Hurt
add DPP diversification to naive dense retrievalNaive RAG · Average of 3 datasets (Nv-embed-v2)F1: 60.62 -> 60.93No clear effect
plug G3RAG in as retriever of iterative frameworksIterDRAG and IterRetgen · Average of 3 datasets (Nv-embed-v2)F1: IterDRAG 62.54 -> 66.26; IterRetgen 64.25 -> 68.42Improved
reduce HippoRAG2 seed count from 5 to 2 (ablation)HippoRAG2 · Average of 3 datasets (Nv-embed-v2)F1: 64.49 -> 63.21Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Baselines: dense retrievers Graph RAG

Builds on: Graph RAG

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy