Source-linked AI summary
Asymmetric Dynamic Routing: Balancing Reasoning Depth and Computational Efficiency in Hypergraph RAG
Qi Sun, Yijia Zhang, Xingliang Hou, Caibo Li, Qiang Li, Yu Guo
TL;DR
Structure-based RAG commonly uses static traversal despite variation in query complexity, creating a mismatch for simple and complex tasks. ADR routes queries by intent among three asymmetric operators; across five domains, it maintains strong reasoning performance while reducing prompt tokens by 48.7% and latency by 45.3%.
Problem
Structure-based RAG systems use uniform subgraph expansion and traversal depth regardless of query complexity, although simple and complex queries have different retrieval needs.
Source: §1 · p. 1Method
ADR is a single-pass, intent-guided framework that adapts retrieval entry level, traversal direction, and expansion scope through localized fact anchoring, bottom-up adjacency diffusion, and top-down insight grounding.
Source: §1 · p. 1Results
Across five domains, ADR maintains strong reasoning performance while reducing prompt tokens by 48.7% and latency by 45.3% relative to Complex-only.
Source: §1 · p. 1, §4.3 · p. 4Takeaways & Limitations
Across five domains, query-adaptive routing maintains strong reasoning performance while reducing unnecessary traversal and context overhead.
Source: §6 · p. 4For Complex queries in structurally sparse domains, limited graph connectivity may constrain Top-Down Insight Grounding; the dual-hypergraph framework also incurs upfront indexing costs.
Source: §5 · p. 4
Abstract
from arXiv · showhide
While graph-based and hypergraph-based Retrieval-Augmented Generation (RAG) significantly mitigate hallucinations in Large Language Models (LLMs), existing structure-based RAG systems typically adopt static traversal strategies regardless of the query complexity. We identify this ``static retrieval fallacy'' as a primary source of computational redundancy for simple queries and cognitive context gaps for complex reasoning tasks. To balance reasoning quality and inference efficiency, we propose Asymmetric Dynamic Routing (ADR), an intent-conditioned retrieval framework operating over hierarchical knowledge graphs. ADR employs a lightweight structured classifier to dynamically dispatch queries among three asymmetric topological traversal operators: localized fact anchoring, bottom-up adjacency diffusion, and top-down insight grounding, which collectively enable bidirectional information flow across hierarchical knowledge layers. Extensive empirical evaluations across five domain-specific corpora demonstrate that ADR maintains strong reasoning performance while reducing prompt token consumption by up to 48.7\% and end-to-end query latency by 45.3\%, yielding a favorable quality--efficiency trade-off for query-adaptive Hypergraph RAG.
1. INTRODUCTION
Structured RAG uses graph and hypergraph representations to support multi-hop synthesis, but uniform traversal can mismatch query complexity. ADR proposes single-pass, intent-guided routing among asymmetric operators; evaluations across five domains report strong reasoning performance with reduced prompt-token consumption and latency.
1. INTRODUCTION
- Uniform subgraph expansion and traversal depth ignore query complexity, a mismatch the paper calls the Static Retrieval Fallacy.Structured graph and hypergraph representations provide multi-hop deductive pathways for synthesis across disparate documents.Source: §1 · p. 1
- Over-retrieval on simple factual queries can add irrelevant context, degrade signal-to-noise ratio, increase prefill memory, and may induce associative hallucinations.Source: §1 · p. 1
- Under-retrieval for abstract cross-document inquiries can leave evidence fragmented and reasoning superficial instead of mechanistic.Localized expansion may fail to bridge disparate abstraction layers.Source: §1 · p. 1
- Reflection-driven methods such as Self-RAG, IR-CoT, and CogRAG add adaptability through multiturn loops but incur multi-second latency overheads.Source: §1 · p. 1
- ADR is a single-pass, intent-guided framework that adjusts retrieval entry level, traversal direction, and expansion scope over a hierarchical knowledge substrate.Its three operators are localized fact anchoring, bottom-up adjacency diffusion, and top-down insight grounding.Source: §1 · p. 1
- Across five domains, ADR maintains strong reasoning performance while substantially reducing prompt-token consumption and latency.The framework is presented as query-adaptive retrieval across heterogeneous traversal requirements.Source: §1 · p. 1
2. RELATED WORK
Prior work organizes knowledge into graphs and hypergraphs, while adaptive retrieval methods vary retrieval frequency or reasoning depth. The related work notes that topological traversal is not explicitly adapted to heterogeneous query requirements.
2. RELATED WORK
- GraphRAG [4] uses LLM-extracted knowledge graphs for hierarchical community summaries, LightRAG dual-level retrieval, GNN-RAG graph neural networks for reasoning paths, and Hyper-RAG hypergraphs to capture pairwise.The passage describes a shift from flat chunk retrieval to topological indexing.Source: §2 · p. 1
- Despite topological advances, retrieval procedures remain largely predefined, potentially over-expanding simple queries and under-covering complex ones.Source: §2 · p. 2
- Adaptive methods select retrieval frequency, trigger retrieval, or correct retrieval quality, but do not explicitly adapt topological traversal to heterogeneous query requirements.The passage also notes that agentic and iterative loops may propagate generation errors and introduce significant temporal latency.Source: §2 · p. 2
3. PROPOSED METHODOLOGY
ADR classifies queries by retrieval and reasoning requirements, then routes them among localized factual retrieval, bottom-up diffusion, and top-down insight grounding. It operates over a hierarchical dual-hypergraph linking source-grounded facts with higher-level semantic structures.
3.1. Hierarchical Dual-Hypergraph Representation
- The dual-hypergraph pairs source-grounded factual entities and relations in HK with abstract structures and semantic associations in HD, linked through higher-order factual hyperedges.Pairwise factual edges remain within HK; higher-order factual hyperedges map to insight vertices, while deep insight edges connect those vertices and carry synthesized narrative insights.Source: §3.1 · p. 2
3.2. Intent Routing and Conservative Fallback
- An LLM-based classifier assigns queries to Simple, Moderate, or Complex categories, accepting predictions at confidence γ ≥ 0.75 and routing lower-confidence predictions to Moderate.Moderate provides an intermediate retrieval scope intended to reduce the risk of either extreme while maintaining bounded retrieval cost.Source: §3.2 · p. 2
3.3.1. Operator 1: Localized Fact Anchoring (ˆy = 1)
- Simple queries retrieve relevant factual entities and their directly associated factual edges in HK, terminating without neighborhood expansion or cross-layer traversal.The resulting granular factual content is formatted as context for generation.Source: §3.3.1 · p. 2
3.3.2. Operator 2: Bottom-Up Adjacency Diffusion (ˆy = 2)
- Moderate queries map higher-order factual edges to insight vertices, then expand once through adjacent deep insight edges in HD.The retrieval context combines the granular factual content with synthesized narrative insights, preserving factual grounding while extending semantic context.Source: §3.3.2 · p. 2
3.3.3. Operator 3: Top-Down Insight Grounding (ˆy = 3)
- Complex queries retrieve relevant deep insight edges from HD, then use inverse cross-layer mapping to obtain supporting factual evidence in HK.The resulting context combines synthesized narrative insights with granular factual content.Source: §3.3.3 · p. 3
4. EXPERIMENTAL EVALUATION
Across five domain settings, the evaluation compares ADR with representative baselines and fixed routing strategies. It examines overall performance, efficiency, routing-threshold sensitivity, and a qualitative reasoning pathway.
4.1. Experimental Configuration
- Five datasets span cross-domain sparse, intra-domain sparse, and intra-domain dense settings; six representative baselines are evaluated with GPT-4o-mini and a five-dimension 0–100 judge.The judge dimensions are Comprehensiveness, Diversity, Empowerment, Logicality, and Readability.Source: §4.1 · p. 3
4.2. Overall Performance Comparison
- ADR scores 83.91 on Mix and 86.65 on Pathology, versus DHI’s 83.18 and 85.78, with similar improvements on CS, Agriculture, and Neurology.Table 2 reports performance across five datasets on a 0–100 scale, with higher scores better.Source: §4.2 · p. 3, Table 2 · p. 4
4.3. Dynamic Routing Analysis and Computational Efficiency
- Against the best fixed traversal configuration, ADR gains 1.39, 2.11, 1.16, 0.94, and 1.25 points on Mix, CS, Agriculture, Neurology, and Pathology.Complex-only scores 83.13 versus Moderate-only’s 82.14 on CS, while Moderate-only leads Complex-only by 1.39 points on Agriculture and 1.51 on Pathology.Source: Table 3 · p. 4, §4.3 · p. 4
- Against Complex-only, ADR cuts average latency 45.3% (435 to 238 ms) and prompt tokens 48.7% (3,760 to 1,929), while overall score rises from 82.05 to 83.91.The comparison profiles computational efficiency and resource parsimony in Table 4.Source: Table 4 · p. 4, §4.3 · p. 4
4.4. Sensitivity Analysis of the Routing Threshold
- γ = 0.75 balances routing reliability and efficiency; at γ ≥0.85, 68.4% of queries fall back to Moderate and latency rises by 58 ms without significant gains.On CS, lowering γ from 0.75 to 0.50 decreases Logicality by 1.14 points.Source: §4.4 · p. 4
4.5. Qualitative Analysis of Reasoning Pathways
- For a Pathology gliomagenesis query, ADR selects Complex (cy = 0.94), grounds a top-down mechanism in supporting facts, and uses 41.2% fewer prompt tokens than Complex-only.The described mechanism links mutant IDH1 to 2-HG accumulation, dioxygenase inhibition, and further epigenetic remodeling.Source: §4.5 · p. 4
5. DISCUSSION AND LIMITATIONS
ADR has structural and computational scope constraints: sparse graph connectivity can limit retrieved evidence, and constructing its dual-hypergraph framework requires upfront indexing.
5. DISCUSSION AND LIMITATIONS
- Sparse graph connectivity can limit factual support for Complex-query grounding, while constructing HK and HD incurs upfront indexing costs.The authors state that insufficient connectivity may require fallback to dense vector search, trading offline construction costs against online inference efficiency.Source: §5 · p. 4
6. CONCLUSION
ADR addresses static traversal in structure-based RAG by adapting retrieval depth and direction to query complexity. Across five domains, it maintains strong reasoning performance while reducing latency and prompt-token use.
6. CONCLUSION
- Across five domains, query-adaptive ADR maintains strong reasoning performance while reducing latency by 45.3% and prompt-token consumption by 48.7%.ADR uses localized fact anchoring, bottom-up adjacency diffusion, and top-down insight grounding to adapt traversal depth and direction.Source: §6 · p. 4