acceptodds
Under review as a conference paper at ICLR 2027

AI-H^2RAG: Asymmetric Indexing-based Hierarchical Hyper-RAG

Abstract

Structured RAG has progressed from modeling pairwise relations with ordinary graphs to representing multi-way, higher-order relations with hypergraphs. However, existing approaches still model cross-granularity hierarchies only to a limited extent, constraining their ability to integrate global context with local facts. Inspired by hierarchical associative memory in the cortico-hippocampal system, we note that human memory jointly encodes multiple elements to link higher-level representations with concrete details. Partial cues can then rapidly reinstate relevant context and facts, with high-level context guiding local interpretation and low-level details grounding holistic recall. Following this principle, we propose Asymmetric Indexing-based Hierarchical Hyper-RAG (AI-H^2RAG). AI-H^2RAG constructs three-layer nested topic–chunk, chunk–sentence, and sentence–entity hypergraphs, organizing evidence so that retrieval can proceed from topic-level scope to entity-level facts and trace local facts back to their topical context. Its asymmetric index propagates final parent context downward and aggregates raw child evidence upward. Each query is then aligned with the index to progressively narrow the candidate space and expand linked evidence for answer generation. Evaluations span eight document corpora and include five generator backbones and three evidence-depth stages. AI-H^2RAG achieves the highest average answer quality across the evaluated settings, with relative gains of up to 7.6% over the strongest baseline. Its online retrieval latency is lower than that of the strongest structured-retrieval baseline, showing that hierarchical higher-order retrieval improves answer quality without sacrificing efficiency.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Retrieval Augmented Generation(RAG), Hypergraph, Asymmetric Indexing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sPxAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To address the shortcomings of existing hypergraph-based RAGs in modeling multi-level relationships, this paper proposes an Asymmetric Indexing-based Hierarchical Hyper-RAG. This method first constructs a three-layer nested hypergraph, then uses Asymmetric Indexing to fuse high-level context with low-level factual information, and finally retrieves and expands relevant evidence progressively through Query Alignment and Coarse-to-Fine Retrieval. Extensive experiments demonstrate the improved answer quality and high online retrieval efficiency of this proposed method.

Critical strengths
  1. This paper constructs a three-layer nested hypergraph and integrates Asymmetric Indexing, Query Alignment, and Coarse-to-Fine Evidence Retrieval into a unified framework, enabling retrieval to simultaneously leverage high-level semantic context and fine-grained factual information.

  2. The paper evaluates AI−H2RAGAI-H^2RAG across eight document corpora, five LLM backbones, and three different evidence depths, demonstrating consistent improvements over the evaluated baselines.

Critical weaknesses
  1. Asymmetric Indexing follows the existing principles of Hierarchical Message Passing [1][2], and its information fusion mechanism is mainly a simple similarity-weighted aggregation. Furthermore, the ablation experiment only compared Asymmetric Indexing with the case without cross-layer propagation, without demonstrating its advantages over other information propagation strategies.
  2. Cue-activated hierarchical evidence recall mainly includes Topic Activation, Coarse-to-Fine Retrieval, Multi-granularity Sentence Reranking, and Entity-mediated Evidence Expansion. The core ideas of these operations have appeared in previous studies [3][4][5]. However, they have not been compared with AI−H2RAGAI-H^2RAG.
  3. Page 2, Figure 1: “Sentences Representaton” is misspelled; it should be “Sentence Representation”.
  4. L272: “We develop a evidence recall” has an incorrect article; it should be “an”.
  5. L294: “We uses these signals” has an agreement error.
  6. L436, L453, L510: “Coarse-to-fine evidence retrieval” is misspelled.

[1] Teng, Zhiyang, and Yue Zhang. "Head-lexicalized bidirectional tree lstms." Transactions of the Association for Computational Linguistics 5 (2017): 163-177.

[2] Zhong, Zhiqiang, Cheng-Te Li, and Jun Pang. "Hierarchical message-passing graph neural networks: Z. Zhong et al." Data Mining and Knowledge Discovery 37.1 (2023): 381-408.

[3] Tao, Wenyu, et al. "Treerag: Unleashing the power of hierarchical storage for enhanced knowledge retrieval in long documents." Findings of the Association for Computational Linguistics: ACL 2025. 2025.

[4] Huang, Haoyu, et al. "Retrieval-Augmented Generation with Hierarchical Knowledge." EMNLP (Findings). 2025.

[5] Zhu, Xiangrong, et al. "Knowledge graph-guided retrieval augmented generation." Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Questions for authors
  1. Could you further elaborate on the key differences between Asymmetric Indexing, Cue-activated hierarchical evidence recall, and existing methods, as well as the innovations of this paper?
  2. "w/o Coarse-to-Fine Evidence Retrieval". Could you conduct a more detailed ablation experiment by removing the following three modules respectively: Topic-scoped Chunk Recall, Sentence Reranking, Entity-mediated Evidence Expansion?
  3. Could you add baselines to compare with other information propagation and retrieval methods, further demonstrating the effectiveness of Asymmetric Indexing and Cue-activated hierarchical evidence recall?
3 comments
AuthorsOpenReview ↗

Here's our reply to the reviewers' question 1

Thank you for your thoughtful review. We respond to each point below.

1. This article first introduces the concept of a hypergraph in the graph construction to build the corresponding hypergraph knowledge base, which is a key difference from RARTOR[1], HiRAG[2], and the Hierarchical Lexical Graph[3]. The reason for choosing a hypergraph is that, as shown by Hyper-rag[4] and hypergraphrag[5], a hypergraph can cover higher-order relationships and provide richer semantics. Next, regarding graph construction, unlike PAPTOR and HiRAG, which use GMM to construct upper-level graphs, and different from the Hierarchical Lexical Graph, which constructs upper-level statements by summarizing triples, we build from topic to chunk, sentence, and entity top-down based on the corresponding ownership relations. In this process, the same chunk may belong to multiple topics, the same sentence may belong to multiple paragraphs, and the same entity may belong to multiple sentences. To handle this, hypergraphs are undoubtedly the most suitable choice. Then, in constructing our retrieval network, regarding indexing, we consider that understanding a passage requires a combination of high-level context and concrete facts. Entities also need to be recognized under relevant semantics for better understanding. Therefore, the corresponding asymmetric information propagation is introduced. It is asymmetric because the vectors transmitted upward and downward are different versions. Parent nodes transmit the final representations fused with context downward, and child nodes transmit their original representations upward. For example, in the case of a “document chunk—sentence”: when a sentence → document chunk, the document chunk needs to know the local evidence it contains, so it aggregates the original vectors of each sentence. These vectors provide the content of the sentences themselves. When a document chunk → sentence, a single sentence may omit the object or describe only a local operation, so it receives the final vector of the document chunk. This chunk vector has already integrated the topic, the content of the chunk itself, and the original sentences inside the chunk, providing the sentence with a broader context. For instance, the original sentence reads, 'Adjust the right-side nut, and tighten after removing the gap.' From an upward perspective, this sentence contributes concrete operation steps to the document chunk. From a downward perspective, the document chunk informs the sentence that this nut belongs to the 'ball screw assembly' and the discussion concerns the 'axial gap.' For the same relation, the information transmitted on both sides serves different purposes. This asymmetric design allows local evidence to borrow from the upper-level context while also avoiding the borrowed context from flowing back as local evidence. Next, regarding the query, we did not directly use the original vectors for the corresponding matching. We also updated the query vectors through corresponding information propagation. At the retrieval level, indeed, our method is conceptually similar to RAPTOR, HiRAG, and Hierarchical Lexical Graph, which perform layer-by-layer retrieval starting from the topic or community level. However, this is a natural idea for constructing a hierarchical hypergraph structure. At the same time, the scoring we use for sentence re-ranking aligns with our design and is not the same as the currently existing scoring algorithms.

[1] Feng, Yifan, et al. "Hyper-RAG: combating LLM hallucinations using hypergraph-driven retrieval-augmented generation." Nature Communications 17, 5778 (2026).

[2] Luo, Haoran, et al. "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation." Advances in Neural Information Processing Systems 38 (2025): 152206-152234.

[3] Sarthi, Parth, et al. "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval." International Conference on Learning Representations. 2024.

[4] Huang, Haoyu, et al. "Retrieval-Augmented Generation with Hierarchical Knowledge." Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. 6044-6060.

[5] Ghassel, Abdellah, et al. "Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval." Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Vol. 2. 2025. 4457-4466.

AuthorsOpenReview ↗

Here's our reply to the reviewers' question 2,3

2. Regarding the ablation experiments, what we did was indeed lacking, so in response to the reviewers' questions, we redesigned the corresponding ablation experiments. The experimental results are as follows:

Ablated designComp.Div.Emp.Log.Read.Rel.Avg.Δ (%)
Full Structure (AI-H²RAG)89.13 ± 0.8779.44 ± 0.7781.57 ± 0.8585.67 ± 0.3588.12 ± 0.3092.53 ± 0.4786.08 ± 0.390.0%
w/o Hierarchical Hypergraph Memory83.80 ± 1.3771.90 ± 1.1474.70 ± 0.6980.60 ± 0.4781.10 ± 0.3185.61 ± 0.4579.62 ± 0.57-7.5%
Ordinary Graph87.60 ± 0.7974.80 ± 0.9275.94 ± 0.8982.70 ± 0.3381.70 ± 0.2291.10 ± 0.2882.31 ± 0.43-4.4%
w/o Asymmetric Hypergraph Network85.00 ± 1.2371.70 ± 1.3375.70 ± 0.7182.50 ± 0.2780.20 ± 0.3791.40 ± 0.4581.08 ± 0.67-5.8%
Asymmetric Raw Aggregation85.80 ± 1.3375.70 ± 0.8774.30 ± 0.7781.60 ± 0.4280.42 ± 0.2990.54 ± 0.7881.39 ± 0.84-5.4%
Symmetric Final Aggregation85.60 ± 1.1475.20 ± 1.5677.26 ± 0.7783.04 ± 0.6885.10 ± 0.3591.20 ± 0.5482.90 ± 0.52-3.7%
w/o Query Alignment81.20 ± 0.8771.40 ± 1.4572.50 ± 0.9080.90 ± 0.7081.72 ± 0.6091.20 ± 0.3179.82 ± 0.64-7.3%
w/o Topic Routing85.50 ± 0.9573.20 ± 0.6375.82 ± 0.9982.84 ± 0.5781.34 ± 0.3590.34 ± 0.3181.51 ± 0.46-5.3%
w/o Chunk-level Coarse Retrieval83.80 ± 1.1571.70 ± 1.3674.96 ± 0.7381.70 ± 0.8481.20 ± 0.5589.50 ± 0.7580.48 ± 0.72-6.5%
w/o Sentence Reranking84.50 ± 1.6774.30 ± 0.9175.16 ± 0.7181.94 ± 0.4580.90 ± 0.4790.60 ± 0.3981.23 ± 0.44-5.6%
w/o Multi-hop Retrieval85.20 ± 0.8473.90 ± 1.2775.30 ± 0.5483.00 ± 0.7084.26 ± 0.4390.80 ± 0.3782.08 ± 0.45-4.6%

This time, we conducted ablation experiments on 30 questions in each of stages 1, 2, and 3 for different groups. First, on the re-layered graph structure, including w/o Hierarchical Hypergraph Memory and Ordinary Graph. Then, on asymmetric retrieval, including w/o Asymmetric Hypergraph Network, Asymmetric Raw Aggregation, and Symmetric Final Aggregation. Finally, on the query, including w/o Query Alignment, w/o Topic Routing, w/o Chunk-level Coarse Retrieval, w/o Sentence Reranking, and w/o Multi-hop Retrieval.

3. This time we compared different levels of RAG on the commonly used public datasets HotpotQA, 2WikiMultiHopQA, and MuSiQue, and the experimental results are as follows:

Methodhotpotqa F1 (%)hotpotqa EM (%)hotpotqa Recall@5 (%)2wikimultihopqa F1 (%)2wikimultihopqa EM (%)2wikimultihopqa Recall@5 (%)musique F1 (%)musique EM (%)musique Recall@5 (%)
Ours72.7660.8085.4060.0054.7074.2145.0435.9054.20
HyperGraphRAG70.3058.5077.4554.4949.4062.2838.5828.5045.67
HiRAG58.5849.1081.4047.4941.7072.0033.3326.0049.20
HGRAG72.1060.6086.6559.2653.8073.2540.4032.4056.39
RAPTOR67.1255.9083.7053.5048.5070.9533.2326.1051.91

It can be seen that our method achieved good results in both F1 score and EM score, while the recall rate on HotpotQA and MuSiQue is slightly lower than HGRAG.

Thank the authors for the responses. Some of my concerns have been addressed. After reading other peer reviews and the responses, I will maintain the original score.

Reviewer_FFvdOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose AI-H2RAG, a hierarchical hypergraph RAG framework that organizes a document corpus into topics, chunks, sentences, and entities to combine global context with local facts. The framework consists of three components: (1) nested topic-chunk, chunk-sentence, and sentence-entity hypergraphs, where topics are Leiden communities over chunk similarity with fuzzy chunk memberships; (2) a training-free asymmetric index that updates each chunk, sentence, and entity embedding with the final representations of its parents and, where present, the raw representations of its children; and (3) a coarse-to-fine retrieval that activates topics, recalls chunks, reranks sentences, and expands to sentences that share entities. The authors evaluate on eight corpora (five machine-tool and three general-domain), with questions generated and scored by GPT-4o-mini on six rubric criteria. They report that AI-H2RAG ranks first under all five generator backbones and reaches relative gains of up to 7.6% over the strongest baseline, which they claim shows that hierarchical higher-order retrieval improves answer quality without sacrificing efficiency.

Critical strengths

S1. The authors tackle an important problem in structured RAG, namely combining global document context with local facts in a single retrieval index. The index is training-free and its cross-level update rules are written out in full (Section 3.2, Appendix D), so it can be applied on top of an off-the-shelf encoder such as BGE-M3.

S2. The authors complement the LLM-judge evaluation with a blinded expert study and a gold-evidence recall analysis (Appendices K and L), both of which favor AI-H2RAG.

Critical weaknesses

W1. Insufficient novelty and justification of key design choices

The novelty over existing hierarchical and hypergraph-based RAG methods is unclear, and the paper does not sufficiently justify its main design choices. This is important because the hierarchical hypergraph, asymmetric indexing, and cue-activated retrieval are presented as the main contributions.

W1.1. The proposed method appears to combine existing ideas without clearly explaining what is new. Section 1 claims that existing RAG methods do not explicitly capture hierarchical document structure, yet the paper neither cites nor discusses closely related work that already organizes evidence across levels. RAPTOR [1] builds and retrieves from a hierarchy of summarized chunks, HiRAG [2] uses hierarchical knowledge in both indexing and retrieval, and HGRAG [3] combines different levels of information through hypergraph retrieval for multi-hop QA. In particular, HGRAG [3] and HyperGraphRAG [4] already use text units (passages or extracted knowledge fragments) as hyperedges over their entities, the same construction as the sentence-entity layer of AI-H2RAG. Likewise, the cue-activated retrieval, which moves from topic-level activation to entity-level expansion, is conceptually similar to the dual-level retrieval of LightRAG, which separates high-level retrieval of broad topics from low-level retrieval of specific entities and is already a baseline in the paper. The authors should explain what is new beyond combining these existing approaches, particularly how the proposed hierarchy, asymmetric indexing, and retrieval process differ from prior work.

W1.2. The need for a hypergraph is unclear. Each hyperedge groups the children of a topic, chunk, or sentence (Section 3.1). The method then aggregates their embeddings using membership-weighted averages and combines them using similarity-based weights (Equations 4 to 6). These relationships and computations could also be represented using a weighted parent-child graph. The authors should explain what benefit the hypergraph provides beyond a standard hierarchical graph.

W1.3. The asymmetric propagation rule is not sufficiently justified. The authors argue that passing final parent embeddings downward and raw child embeddings upward prevents contextual information from being reused as factual evidence (Section 3.2). However, they do not explain why preventing this reuse should improve retrieval, or show that symmetric propagation actually harms it.

[1] Sarthi, Parth, et al. "Raptor: Recursive abstractive processing for tree-organized retrieval." International Conference on Learning Representations. Vol. 2024. 2024.

[2] Huang, Haoyu, et al. "Retrieval-Augmented Generation with Hierarchical Knowledge." EMNLP (Findings). 2025.

[3] Wang, Changjian, et al. "Cross-granularity hypergraph retrieval-augmented generation for multi-hop question answering." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 39. 2026.

[4] Luo, Haoran, et al. "Hypergraphrag: Retrieval-augmented generation via hypergraph-structured knowledge representation." Advances in Neural Information Processing Systems 38 (2026): 152206-152234.

W2. Insufficient experimental validation

The experiments do not sufficiently establish whether AI-H2RAG produces more accurate answers than strong existing methods or whether its key design choices explain the improvements. This is important given the paper's claims about answer quality and efficiency.

W2.1. The main evaluation relies on synthetic questions and metrics that do not directly measure answer correctness. GPT-4o-mini generates questions from sampled document chunks and evaluates answers using six criteria (Appendices H–I). Although Appendix L includes human-annotated evidence for 120 questions, the paper does not report systematic validation of answerability or required evidence connections across the full evaluation set. Moreover, none of the six criteria directly measures factual agreement with the reference evidence, while Diversity rewards additional content. The no-retrieval LLM even outperforms dense RAG under three of five generators (e.g., 80.0 vs. 75.1 with DeepSeek-V3.2, Figure 5), further motivating a direct correctness evaluation. The expert correctness and evidence recall studies (Appendices K–L) cover only 60 and 120 MechanicalCorp questions, respectively, comparing only against Hyper-RAG. The main comparisons also do not establish whether generator context budgets are matched. The authors should evaluate on public multi-hop QA benchmarks with gold answers, such as HotpotQA, 2WikiMultiHopQA, and MuSiQue, reporting answer accuracy and retrieval recall under matched context budgets.

W2.2. The baseline comparison omits important competitors. RAPTOR [1], HiRAG [2], and HGRAG [3] directly address hierarchical or cross-granularity retrieval. Other relevant methods include HyperGraphRAG [4] and PropRAG [5], which reports higher Recall@5 and answer F1 than HippoRAG 2 on HotpotQA, 2WikiMultiHopQA, and MuSiQue. The authors should include these methods as baselines.

W2.3. The ablations are too coarse to identify which design choices drive the improvements. Removing the hierarchical hypergraph also removes the hierarchy, replacing it with a flat dense chunk index (Section 4.4), so the benefit of hyperedges over a standard hierarchical graph remains unclear. Similarly, removing asymmetric indexing disables all cross-level propagation, while the coarse-to-fine retrieval ablation removes topic activation, sentence reranking, and entity-based expansion together. These experiments do not isolate the contribution of each design choice. Moreover, all ablations use only Stage 1 questions, despite the paper's focus on multi-hop retrieval. The authors should provide more fine-grained ablations comparing hierarchical graphs with hypergraphs, symmetric with asymmetric propagation, and individual retrieval components, including on Stages 2 and 3.

W2.4. The efficiency evaluation excludes a major source of cost. The abstract claims improved answer quality "without sacrificing efficiency," but Section 4.5 measures only online retrieval and context construction, excluding index construction. Building the index requires LLM-based sentence and entity extraction, embeddings at multiple levels, and topic clustering over corpora containing millions of words (Appendices A to C, Table I.1). The method is also slower online than dense RAG and LightRAG (Figure 9). The authors should report indexing time, LLM cost, and index size, and qualify the efficiency claim.

[5] Wang, Jingjin, and Jiawei Han. "Proprag: Guiding retrieval with beam search over proposition paths." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Questions for authors

Please refer to W1 and W2.

5 comments
AuthorsOpenReview ↗

Here's our reply to the reviewers' w1,1 question

Thank you for your thoughtful review. We respond to each point below.

W1.1. We drew on Hyper-RAG [1] and HyperGraphRAG [2] for the idea of using hypergraphs to construct a knowledge base. Hyper-RAG [1], HyperGraphRAG [2], and HGRAG [3] focus on relationships between chunks and entities, whereas our construction explicitly accounts for the hierarchical organization of documents. RAPTOR [4] and HiRAG [5] use GMMs to construct the upper levels of their hierarchies. Our approach builds the hierarchy top-down from topics to chunks, sentences, and entities based on the membership relations between adjacent levels. A single chunk can belong to multiple topics, a single sentence can belong to multiple passages, and a single entity can appear in multiple sentences. We consider hypergraphs the most suitable representation for these overlapping memberships.

In constructing the retrieval network, we designed the index around the observation that fully interpreting a passage requires both higher-level context and specific facts, and that entities are easier to interpret in their semantic context. This motivated our asymmetric information propagation scheme. The asymmetry comes from using different versions of the embeddings in the two directions. Parent nodes pass their final, context-enriched representations downward, while child nodes pass their own raw representations upward. Take the “chunk–sentence” relationship as an example. In the “sentence → chunk” direction, the chunk aggregates the raw embeddings of its sentences to capture the local evidence they contain. These embeddings represent the content of the sentences themselves. In the “chunk → sentence” direction, a sentence receives the final embedding of its parent chunk because the sentence may omit the object of an instruction or describe only one step of a larger procedure. The chunk’s final embedding combines topic-level context, the chunk’s own content, and the raw embeddings of its sentences, giving each sentence access to broader context. For example, consider the sentence “Adjust the nut on the right, then lock it in place once the clearance has been eliminated.” In the upward direction, the sentence contributes a specific procedural step to the chunk. In the downward direction, the chunk provides the context that the nut is part of a ball screw assembly and that the clearance is axial. The two directions thus serve different purposes along the same relationship. This asymmetric design gives local evidence access to higher-level context and prevents that inherited context from flowing back to the parent as local evidence.

Regarding the query, we did not directly use the original vectors for the corresponding matching. Instead, we also updated the query vectors through the propagation of information. At the retrieval level, our method indeed follows a similar idea to RAPTOR[4] and HiRAG[5], which perform retrieval layer by layer starting from the topic. However, this is a very normal approach when constructing hierarchical hypergraph structures. At the same time, the scoring we use for sentence re-ranking aligns with our design, and it is different from the scoring algorithms currently in use.

[1] Feng, Yifan, et al. "Hyper-RAG: combating LLM hallucinations using hypergraph-driven retrieval-augmented generation." Nature Communications 17, 5778 (2026).

[2] Luo, Haoran, et al. "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation." Advances in Neural Information Processing Systems 38 (2025): 152206-152234.

[3] Wang, Changjian, et al. "Cross-Granularity Hypergraph Retrieval-Augmented Generation for Multi-hop Question Answering." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 39. 2026.

[4] Sarthi, Parth, et al. "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval." International Conference on Learning Representations. 2024.

[5] Huang, Haoyu, et al. "Retrieval-Augmented Generation with Hierarchical Knowledge." Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. 6044-6060.

[6] Ghassel, Abdellah, et al. "Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval." Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Vol. 2. 2025. 4457-4466.

AuthorsOpenReview ↗

Here's our reply to the reviewers' w1.2 ,w1.3 questions

W1.2. Regarding the necessity of hypergraphs, we know from Hyper-rag[1] and hypergraphrag[2] that hypergraphs can capture higher-order relationships and provide richer semantics. Furthermore, according to the construction described in w1.1, situations arise where the same chunk belongs to multiple topics, the same sentence belongs to multiple paragraphs, and the same entity belongs to multiple sentences. To handle such situations, there is no doubt that hypergraphs are the most suitable, because hypergraphs can treat an entire set of members as a single hyperedge with identity, and the same node can appear in multiple hyperedges simultaneously. However, we do have shortcomings; we did not explain in the ablation experiments whether hypergraphs truly offer benefits compared to ordinary graphs. Subsequent ablation experiments have addressed this, and refer to the experimental section below for responses regarding the ablation experiments.

W1.3. As mentioned in the answer to W1.1, we consider that interpreting a passage requires a combination of high-level context and specific facts in order to be truly understood, and entities also need to be recognized under the corresponding semantics to be better understood. Therefore, there is corresponding asymmetric information propagation. It is asymmetric because the vectors read during upward transmission and downward transmission are different versions. The parent node transmits downward the final representation that has already fused context, while the child node transmits upward its original representation. This avoids the situation where if the child node transmits the final vector representation, the information obtained by the parent node would flow back as local evidence to the parent node itself. Of course, regarding this point, our ablation experiments were indeed insufficient, but subsequent ablation experiments have been added. Please refer to the following experimental section for the answers regarding ablation experiments.

AuthorsOpenReview ↗

Here's our reply to the reviewers' w2.1,w2.2,w2.3 questions

W2.1. Our experimental comparison section indeed lacks comparisons on the commonly used public datasets HotpotQA, 2WikiMultiHopQA, and MuSiQue, with different levels of RAG, so we conducted the corresponding experiments, and the results are as follows:

Methodhotpotqa F1 (%)hotpotqa EM (%)hotpotqa Recall@5 (%)2wikimultihopqa F1 (%)2wikimultihopqa EM (%)2wikimultihopqa Recall@5 (%)musique F1 (%)musique EM (%)musique Recall@5 (%)
Ours72.7660.8085.4060.0054.7074.2145.0435.9054.20
HyperGraphRAG70.3058.5077.4554.4949.4062.2838.5828.5045.67
HiRAG58.5849.1081.4047.4941.7072.0033.3326.0049.20
HGRAG72.1060.6086.6559.2653.8073.2540.4032.4056.39
RAPTOR67.1255.9083.7053.5048.5070.9533.2326.1051.91

It can be seen that our method achieved good results in both F1 score and EM score, while the recall rate on HotpotQA and MuSiQue is slightly lower than HGRAG.

W2.2. This time we compared PAPTOR, HiRAG, Hierarchical Lexical Graph[6], HyperGraph[2] on commonly used public datasets HotpotQA, 2WikiMultiHopQA, and MuSiQue, and the results are shown in Table w2.1.

W2.3. Regarding the ablation experiments, what we did was indeed lacking, so in response to the reviewers' questions, we redesigned the corresponding ablation experiments. The experimental results are as follows:

Ablated designComp.Div.Emp.Log.Read.Rel.Avg.Δ (%)
Full Structure (AI-H²RAG)89.13 ± 0.8779.44 ± 0.7781.57 ± 0.8585.67 ± 0.3588.12 ± 0.3092.53 ± 0.4786.08 ± 0.390.0%
w/o Hierarchical Hypergraph Memory83.80 ± 1.3771.90 ± 1.1474.70 ± 0.6980.60 ± 0.4781.10 ± 0.3185.61 ± 0.4579.62 ± 0.57-7.5%
Ordinary Graph87.60 ± 0.7974.80 ± 0.9275.94 ± 0.8982.70 ± 0.3381.70 ± 0.2291.10 ± 0.2882.31 ± 0.43-4.4%
w/o Asymmetric Hypergraph Network85.00 ± 1.2371.70 ± 1.3375.70 ± 0.7182.50 ± 0.2780.20 ± 0.3791.40 ± 0.4581.08 ± 0.67-5.8%
Asymmetric Raw Aggregation85.80 ± 1.3375.70 ± 0.8774.30 ± 0.7781.60 ± 0.4280.42 ± 0.2990.54 ± 0.7881.39 ± 0.84-5.4%
Symmetric Final Aggregation85.60 ± 1.1475.20 ± 1.5677.26 ± 0.7783.04 ± 0.6885.10 ± 0.3591.20 ± 0.5482.90 ± 0.52-3.7%
w/o Query Alignment81.20 ± 0.8771.40 ± 1.4572.50 ± 0.9080.90 ± 0.7081.72 ± 0.6091.20 ± 0.3179.82 ± 0.64-7.3%
w/o Topic Routing85.50 ± 0.9573.20 ± 0.6375.82 ± 0.9982.84 ± 0.5781.34 ± 0.3590.34 ± 0.3181.51 ± 0.46-5.3%
w/o Chunk-level Coarse Retrieval83.80 ± 1.1571.70 ± 1.3674.96 ± 0.7381.70 ± 0.8481.20 ± 0.5589.50 ± 0.7580.48 ± 0.72-6.5%
w/o Sentence Reranking84.50 ± 1.6774.30 ± 0.9175.16 ± 0.7181.94 ± 0.4580.90 ± 0.4790.60 ± 0.3981.23 ± 0.44-5.6%
w/o Multi-hop Retrieval85.20 ± 0.8473.90 ± 1.2775.30 ± 0.5483.00 ± 0.7084.26 ± 0.4390.80 ± 0.3782.08 ± 0.45-4.6%

This time, we conducted ablation experiments on 30 questions in each of stages 1, 2, and 3 for different groups. First, on the re-layered graph structure, including w/o Hierarchical Hypergraph Memory and Ordinary Graph. Then, on asymmetric retrieval, including w/o Asymmetric Hypergraph Network, Asymmetric Raw Aggregation, and Symmetric Final Aggregation. Finally, on the query, including w/o Query Alignment, w/o Topic Routing, w/o Chunk-level Coarse Retrieval, w/o Sentence Reranking, and w/o Multi-hop Retrieval.

Reviewer_MAXmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AI-H2RAG, a RAG framework that organizes topics, chunks, sentences, and entities into hierarchical hypergraphs. It combines asymmetric information propagation with coarse-to-fine retrieval to integrate contextual and factual evidence. Experiments across multiple datasets and generator backbones report consistent improvements over existing RAG baselines.

Critical strengths

S1. The paper is generally well organized, with a clear motivation and a coherent framework.

S2. The experimental evaluation covers multiple datasets, baselines, and model settings, with generally consistent improvements.

Critical weaknesses

W1. The methodological novelty appears somewhat limited, as several key ideas overlap with existing hierarchical retrieval methods.

W2. The specific advantages of asymmetric propagation and nested hypergraph modeling are not sufficiently established.

W3. The evaluation provides limited insight into the overall computational cost and scalability of the proposed indexing framework.

Questions for authors

D1. The paper could better distinguish its contributions from related hierarchical retrieval approaches, such as RAPTOR, HiRAG, and Hierarchical Lexical Graph.

D2. The ablation in Section 4.4 demonstrates the benefit of cross-level propagation but does not isolate the specific contribution of asymmetry. Similarly, the necessity of using hypergraphs rather than simpler hierarchical representations remains unclear.

D3. Section 4.5 reports online retrieval latency but excludes offline index construction. A more complete characterization of indexing overhead would strengthen the efficiency claims.

3 comments
AuthorsOpenReview ↗

Here's our reply to the reviewers' D1 question

D1. This paper first introduces the concept of hypergraphs in the graph construction to build the corresponding hypergraph knowledge base. This is a key aspect compared with RARTOR [1], HiRAG [2], and Hierarchical Lexical Graph [3]. The reason for choosing hypergraphs is that, as demonstrated by Hyper-rag [4] and hypergraphrag [5], hypergraphs can cover higher-order relationships and provide richer semantics. Secondly, in terms of graph construction, unlike PAPTOR and HiRAG which use GMM for upper-layer graph construction, or Hierarchical Lexical Graph which builds the upper-layer statement by summarizing triples, we construct it top-down from topics to chunks, sentences, and entities according to their respective affiliation relationships. During this process, the same chunk may belong to multiple topics, the same sentence may belong to multiple paragraphs, and the same entity may belong to multiple sentences. Dealing with this situation, without a doubt, hypergraphs are most suitable. Then comes the construction of our retrieval network. In terms of indexing, we consider that understanding a passage requires the combination of a higher-level context and specific facts. Entities also need to be understood within the corresponding semantics. Therefore, there is an asymmetric information propagation. The propagation is asymmetric because the vectors read during upward and downward transmission are different versions. The parent node passes down a final representation already fused with context, while the child node passes up its original representation. Taking the "document chunk—sentence" as an example: from sentence → document chunk, the document chunk needs to know the local evidence it contains, so it aggregates the original vectors of each sentence. These vectors provide the content of the sentences themselves. From document chunk → sentence, a sentence may omit objects or only describe a local operation, so it receives the final vector of the document chunk. This chunk vector, already fused with the topic, content of the chunk, and original sentences within the chunk, can supplement the sentence with a broader context. For example, the original sentence says, "Adjust the right nut, lock it after eliminating the gap." Looking upward, this sentence contributes specific operational steps to the document chunk. Looking downward, the document chunk informs this sentence that the nut belongs to the "ball screw pair" and concerns the "axial gap." For the same relationship, the information conveyed on each side serves different purposes. This asymmetric design allows local evidence to borrow higher-level context while preventing the borrowed context from flowing back as local evidence. Next, regarding queries, we do not directly use the original vectors for matching; we also update the query vectors through information propagation. In terms of retrieval, our method indeed follows the same idea as RAPTOR, HiRAG, and Hierarchical Lexical Graph, starting retrieval layer by layer from the topic or community level. However, this is a natural approach to constructing a hierarchical hypergraph structure. Moreover, the scoring we use for sentence re-ranking aligns with our design, which differs from the existing scoring algorithms. At the same time, we also added some comparative experiments,results are as follows:

Methodhotpotqa F1 (%)hotpotqa EM (%)hotpotqa Recall@5 (%)2wikimultihopqa F1 (%)2wikimultihopqa EM (%)2wikimultihopqa Recall@5 (%)musique F1 (%)musique EM (%)musique Recall@5 (%)
Ours72.7660.8085.4060.0054.7074.2145.0435.9054.20
HyperGraphRAG70.3058.5077.4554.4949.4062.2838.5828.5045.67
HiRAG58.5849.1081.4047.4941.7072.0033.3326.0049.20
HGRAG72.1060.6086.6559.2653.8073.2540.4032.4056.39
RAPTOR67.1255.9083.7053.5048.5070.9533.2326.1051.91

References follow in Part 2.

AuthorsOpenReview ↗

Here's our reply to the reviewers' D2 question

References for D1

[1] Sarthi, Parth, et al. "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval." International Conference on Learning Representations. 2024.

[2] Huang, Haoyu, et al. "Retrieval-Augmented Generation with Hierarchical Knowledge." Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. 6044-6060.

[3] Ghassel, Abdellah, et al. "Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval." Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Vol. 2. 2025. 4457-4466.

[4] Feng, Yifan, et al. "Hyper-RAG: combating LLM hallucinations using hypergraph-driven retrieval-augmented generation." Nature Communications 17, 5778 (2026).

[5] Luo, Haoran, et al. "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation." Advances in Neural Information Processing Systems 38 (2025): 152206-152234.

[1] Sarthi, Parth, et al. "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval." International Conference on Learning Representations. 2024.

[2] Huang, Haoyu, et al. "Retrieval-Augmented Generation with Hierarchical Knowledge." Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. 6044-6060.

[3] Ghassel, Abdellah, et al. "Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval." Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Vol. 2. 2025. 4457-4466.

[4] Feng, Yifan, et al. "Hyper-RAG: combating LLM hallucinations using hypergraph-driven retrieval-augmented generation." Nature Communications 17, 5778 (2026).

[5] Luo, Haoran, et al. "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation." Advances in Neural Information Processing Systems 38 (2025): 152206-152234.

D2. Regarding the ablation experiments, what we actually did was indeed somewhat lacking, so in response to the issues raised by the reviewers, we have redesigned the corresponding ablation experiments. The experimental results are as follows:

Ablated designComp.Div.Emp.Log.Read.Rel.Avg.Δ (%)
Full Structure (AI-H²RAG)89.13 ± 0.8779.44 ± 0.7781.57 ± 0.8585.67 ± 0.3588.12 ± 0.3092.53 ± 0.4786.08 ± 0.390.0%
w/o Hierarchical Hypergraph Memory83.80 ± 1.3771.90 ± 1.1474.70 ± 0.6980.60 ± 0.4781.10 ± 0.3185.61 ± 0.4579.62 ± 0.57-7.5%
Ordinary Graph87.60 ± 0.7974.80 ± 0.9275.94 ± 0.8982.70 ± 0.3381.70 ± 0.2291.10 ± 0.2882.31 ± 0.43-4.4%
w/o Asymmetric Hypergraph Network85.00 ± 1.2371.70 ± 1.3375.70 ± 0.7182.50 ± 0.2780.20 ± 0.3791.40 ± 0.4581.08 ± 0.67-5.8%
Asymmetric Raw Aggregation85.80 ± 1.3375.70 ± 0.8774.30 ± 0.7781.60 ± 0.4280.42 ± 0.2990.54 ± 0.7881.39 ± 0.84-5.4%
Symmetric Final Aggregation85.60 ± 1.1475.20 ± 1.5677.26 ± 0.7783.04 ± 0.6885.10 ± 0.3591.20 ± 0.5482.90 ± 0.52-3.7%
w/o Query Alignment81.20 ± 0.8771.40 ± 1.4572.50 ± 0.9080.90 ± 0.7081.72 ± 0.6091.20 ± 0.3179.82 ± 0.64-7.3%
w/o Topic Routing85.50 ± 0.9573.20 ± 0.6375.82 ± 0.9982.84 ± 0.5781.34 ± 0.3590.34 ± 0.3181.51 ± 0.46-5.3%
w/o Chunk-level Coarse Retrieval83.80 ± 1.1571.70 ± 1.3674.96 ± 0.7381.70 ± 0.8481.20 ± 0.5589.50 ± 0.7580.48 ± 0.72-6.5%
w/o Sentence Reranking84.50 ± 1.6774.30 ± 0.9175.16 ± 0.7181.94 ± 0.4580.90 ± 0.4790.60 ± 0.3981.23 ± 0.44-5.6%
w/o Multi-hop Retrieval85.20 ± 0.8473.90 ± 1.2775.30 ± 0.5483.00 ± 0.7084.26 ± 0.4390.80 ± 0.3782.08 ± 0.45-4.6%

This time, we conducted ablation experiments on 30 questions in each of stages 1, 2, and 3 for different groups. First, on the re-layered graph structure, including w/o Hierarchical Hypergraph Memory and Ordinary Graph. Then, on asymmetric retrieval, including w/o Asymmetric Hypergraph Network, Asymmetric Raw Aggregation, and Symmetric Final Aggregation. Finally, on the query, including w/o Query Alignment, w/o Topic Routing, w/o Chunk-level Coarse Retrieval, w/o Sentence Reranking, and w/o Multi-hop Retrieval.

AuthorsOpenReview ↗

Here's our reply to the reviewers' D3 question

D3. Regarding retrieval efficiency, we indeed have some shortcomings. Therefore, this time we will display the results in Indexing and Retrieval through three major metrics: LLM Token Cost, LLM API Calls, and Time Cost. Due to considerations of time and comparative cost, we trimmed the corpus when conducting this experiment, reducing it to 61,461 tokens. The results are as follows:

MethodLLM Tokens / IndexingLLM Tokens / RetrievalLLM Calls / IndexingLLM Calls / RetrievalTime (s) / IndexingTime (s) / Retrieval
RAG0.000.000.000.007.410.17
LightRAG798,695.00689.58105.001.00281.911.18
HyperRAG953,591.00448.8798.001.00228.062.38
HiRAG3,477,523.000.00718.000.00674.110.21
RAPTOR82,164.000.00137.000.00235.520.20
Ours482,277.002,372.7795.002.00129.173.67

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.