Source-linked AI summary
ISO-RAG: Isoperimetric Noise Control for Retrieval-Augmented Generation
Siyuan Zhang, Hanchen Wang, Dong Wen, Ying Zhang, Wenjie Zhang
TL;DR
Multi-hop QA exposes weaknesses in flat retrieval and graph-based RAG, particularly incomplete reasoning chains and noisy diffusion. ISO-RAG projects graph nodes into hyperbolic space and applies isoperimetric control to prune misleading connections before localized diffusion. Across benchmarks, it reports consistent gains in retrieval and downstream QA, while results also show that generation can remain a separate bottleneck.
Problem
Flat retrieval struggles to reconstruct the complete reasoning chains needed for multi-hop QA, while graph-based retrieval can suffer from spurious diffusion.
Method
ISO-RAG maps graph-supervised text representations into a hyperbolic Poincaré-ball space and uses node-wise isoperimetric profiles to prune misleading connections before retrieval diffusion.
Results
ISO-RAG delivers nearly 10% average absolute Recall@5 gains and 4.3% Exact Match gains over competitive baselines across multi-hop QA evaluations.
Takeaways & Limitations
Topology-aware control supports efficient, effective graph retrieval for multi-hop reasoning, with consistent absolute improvements over competitive baselines.
Takeaways & Limitations
Even with 100% ground-truth entity retrieval, the LLM can fail, so degraded F1 or EM may reflect generation bottlenecks rather than retrieval failure alone.
Abstract
from arXiv · showhide
Retrieval-Augmented Generation (RAG) mitigates large language models (LLMs) hallucinations, yet conventional dense retrieval struggles with the complex reasoning paths of multi-hop question answering (QA). Graph-based RAG captures multi-step relationships but suffers from severe semantic drift and high online latency due to noisy global graph traversals. Thus, we propose ISO-RAG (ISOperimetric Retrieval-Augmented Generation), a geometry-aware RAG framework. By projecting the underlying knowledge graph into a hyperbolic Poincare ball to precompute node-wise isoperimetric profiles, ISO-RAG prunes spurious edges during retrieval, restricting the search space to a strictly localized subgraph. This topological purification regulates Personalized PageRank (PPR) diffusion driving the retrieval process, ensuring exact and low-latency convergence. Experiments on multi-hop QA benchmarks demonstrate that ISO-RAG outperforms state-of-the-art baselines by average absolute gains of 10.0% in retrieval recall and 4.3% in downstream exact match, achieving a superior accuracy-efficiency trade-off by fundamentally eliminating the latency bottleneck of global traversals. Our source code is available at https://github.com/ZaiizaiZHANG/ISO-RAG.
1 Introduction
RAG improves factuality but remains unreliable for multi-hop QA because flat retrieval and unconstrained graph diffusion can fail to preserve complete reasoning chains. ISO-RAG addresses this with hyperbolic geometry and explicit topological pruning, reporting gains in retrieval and downstream QA performance.
- Multi-hop QA requires connecting evidence scattered across documents, making retrieval quality a dominant bottleneck for reliable answers.
- Flat sparse and dense retrievers often fail to reconstruct the full reasoning chains required by compositional benchmarks.
- Graph-based RAG organizes corpora into interconnected structures and hyperbolic embeddings can capture complex multi-hop dependencies with low structural distortion.
- Unconstrained PPR diffusion over dense graphs causes semantic drift by propagating retrieval signals toward misleading cues and irrelevant noise.
- ISO-RAG forms a candidate subgraph, maps nodes into hyperbolic space, and uses isoperimetric control to sever erroneous branches before diffusion.
- Nearly 10% average absolute Recall@5 gains and 4.3% Exact Match gains over competitive baselines indicate a favorable retrieval-efficiency and QA-performance balance.
2 Related Works
Prior work progresses from flat retrieval toward graph-based and hyperbolic representations for multi-hop evidence aggregation. However, graph systems can still retrieve spurious subgraphs because their diffusion and edge weighting lack explicit topological bounds.
- Sparse and dense retrievers operate in flat search spaces and cannot explicitly model relationships between intermediate entities.
- Graph-based retrieval structures corpora into networks that capture multi-hop dependencies through hierarchical summaries, dual-level structures, or continuous activation spreading.
- These graph-based baselines rely on heuristic edge weighting and unconstrained probability diffusion despite improving recall.
- Without mathematical bounds that prune the search space, graph frameworks retrieve spurious subgraphs and misleading entities before generation.
- Knowledge-graph hierarchy motivates hyperbolic representations, including Poincaré-ball and Lorentz-model approaches.
3 Methodology
ISO-RAG builds a query-conditioned local graph, maps its representations into hyperbolic space, and filters structurally inconsistent edges before deterministic Personalized PageRank. This localized design combines semantic seed routing with geometric pruning to limit diffusion and preserve coherent multi-hop paths.
- ISO-RAG combines query-aware seed routing, local candidate graph construction, hyperbolic isoperimetric filtering, and deterministic Personalized PageRank.
- The retrieval graph represents passages as nodes and connects passages that co-occur in training instances or retrieval contexts.This exposes latent multi-hop dependencies but also introduces noisy shortcuts and hub-like regions.
- The local candidate set unions a dense semantic pool with a k-hop neighborhood expanded from the query’s top-m seed nodes.The induced subgraph reduces the search space while retaining structural connectivity around semantically aligned entry points.
- ISO-RAG maps node representations into the Poincaré ball using an offline-trained projection that captures textual semantics and hierarchical graph structure.The projection is trained with a graph-supervised margin-based triplet loss and a radial depth regularizer.
- The node-wise isoperimetric score compares hyperbolic boundary and internal volumes to identify structural bottlenecks and penalize generic hubs.The internal region is a closed 1-hop neighborhood, while its boundary is a 2-hop shell; the same hyperbolic measure is used for both volumes.
- Filtered edges exclude structurally divergent endpoints, after which deterministic PPR diffuses only on the compact local graph from seed-weighted personalization.This bounds probability mass within the local manifold and enables exact, efficient computation via power iteration.
4 Experiments
ISO-RAG is evaluated on three multi-hop QA benchmarks for retrieval quality, downstream QA, efficiency, and geometric-filtering behavior. It consistently performs strongly across reasoning settings while reducing retrieval latency and showing that learned, topology-aligned filtering drives the gains.
- Overall Performance: ISO-RAG consistently yields the best retrieval and end-to-end QA results across evaluated settings.Retrieval is assessed with Recall and Precision, while downstream generation uses Exact Match and F1 under identical prompting.
- Overall Performance: 88.10 R@5 on 2WikiMultihopQA represents a 10.17-point absolute improvement over LightRAG, while Qwen3-Max reaches 84.7 F1 and 80.7 EM.The gains are reported for a complex-topology benchmark requiring preservation of valid intermediate hops and suppression of spurious diffusion.
- Overall Performance: 84.00 R@10 on MuSiQue gives ISO-RAG an 11.4% relative gain over HippoRAG2, while Qwen-Plus reaches 42.8 F1 and an approximate 25% relative gain.The results are reported under dense distractors and deep compositional reasoning, where precise multi-hop evidence pathways are especially important.
- Efficiency Analysis: ISO-RAG maintains millisecond-level retrieval speeds, including an approximate 8× speedup over HippoRAG2 and more than 25× over HyperbolicRAG on MuSiQue.It ranks as the second-fastest retriever on HotpotQA and MuSiQue, while operating on localized rather than globally traversed graphs.
- Geometric Filtering: 87.28% peak recall and 40.44% peak precision at β = 0.40 show that real_phi outperforms uniform_phi and random_phi, which remain near 71–72%.The ablation evaluates recall, precision, and removed-edge percentage while varying Candidate X, β, and ϕ mapping strategies.
- Geometric Filtering: Topology-aware filtering reaches peak results after pruning 32% of edges, whereas shuffled_phi removes up to 75% and suffers a major performance drop.This comparison separates aligned geometric guidance from indiscriminate edge removal.
- Qualitative Case Study: ISO-RAG can retrieve all ground-truth entities while the language model still fails, showing that degraded QA scores may reflect generation bottlenecks rather than retrieval failure.The qualitative case study describes a “Perfect Retrieval, Failed Generation” example.
5 Conclusion
ISO-RAG imposes strict topological control on graph-based retrieval to mitigate spurious diffusion. Across benchmarks, it produces consistent absolute improvements over competitive baselines for downstream question answering.
- ISO-RAG imposes strict topological control to mitigate spurious diffusion during graph-based retrieval.It maps the graph into hyperbolic space and trains a geometry-aware encoder for controlled retrieval.
- Across benchmarks, ISO-RAG yields consistent absolute improvements over competitive baselines.The reported gains support a balance between retrieval efficiency and downstream QA performance.