acceptodds
Under review as a conference paper at ICLR 2027

HyperReCo: Retrieving and Connecting Evidence with Hypergraph Neural Networks for LLM Multi-hop Reasoning

Abstract

Large language models (LLMs) have shown strong capabilities, with retrieval-augmented generation (RAG) supporting complex multi-hop reasoning by retrieving evidence distributed across documents. Graph-based approaches exploit connections among evidence, and hypergraph-based retrieval further preserves higher-order entity associations within documents and connects documents through shared entities. However, existing hypergraph retrievers often rely on predefined structural expansion or diffusion, which may miss query-dependent interactions needed to identify relevant evidence. They also leave connections among retrieved evidence implicit, requiring LLMs to reconstruct these connections before reasoning. Therefore, we propose HyperReCo, a framework for retrieving and connecting evidence with a hypergraph neural network (HyperGNN). We represent each document as a hyperedge over its extracted entities, with shared entities connecting the hyperedges. Through hypergraph message passing with joint supervision over documents and entities, the HyperGNN learns query-dependent interactions to retrieve complementary evidence. We further introduce Gradient-Guided Hyper-Path Decoding (GGHD), which uses gradient attribution to interpret the learned interactions and translate them into explicit hyper-paths that help LLMs combine complementary facts for multi-hop reasoning. Experiments on six benchmarks show that HyperReCo achieves the best retrieval performance among the compared methods on all three multi-hop QA datasets, together with strong downstream QA performance. Case studies and further analyses demonstrate the utility of decoded hyper-paths for connecting retrieved evidence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
HyperGraph Neural Network, Retrieval-augmented Generation, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QK1COpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HyperReCo, a hypergraph-based framework for multi-hop retrieval-augmented generation. It introduces a query-conditioned HyperGNN to learn document and entity relevance through hypergraph message passing, and Gradient-Guided Hyper-Path Decoding (GGHD) to construct explicit evidence connections for LLM reasoning. Experiments on six benchmarks demonstrate strong retrieval and downstream QA performance, with ablation studies and cross-domain evaluations supporting the effectiveness of the framework.

Critical strengths
  1. Strong and consistent empirical performance. HyperReCo achieves consistent improvements in both retrieval recall and downstream QA accuracy across three multi-hop benchmarks. The gains over HGRAG are substantial, and the additional experiments with different embedding models further support the robustness of the approach. These results provide convincing evidence of the framework's practical effectiveness.

  2. Effective integration of retrieval and evidence connection. The paper addresses a meaningful limitation of multi-hop RAG: retrieving the correct documents does not guarantee that an LLM can connect their evidence. The path ablation shows that explicit evidence connections improve mean EM from 55.0 to 57.6 under the same retrieved documents. This provides useful evidence for the value of the proposed retrieval-and-connection framework.

Critical weaknesses
  1. Limited methodological novelty relative to prior work. The main components are closely related to existing approaches: HGRAG adopts a similar document–entity hypergraph representation, while GFM-RAG and HyperRAG have explored learned structural retrieval and evidence-path construction. Although HyperReCo combines these ideas effectively, the paper does not sufficiently establish what new capabilities its HyperGNN provides beyond existing graph-based approaches. This limits the significance of the methodological contribution.

  2. Insufficient evidence for the distinctive contribution of GGHD. While explicit paths clearly improve QA performance, GGHD only achieves a 0.6–0.7 point mean EM improvement over the simpler S-forest and G-paths alternatives. The statistical significance of these gains is not established, and the faithfulness of the gradient-based connection scores remains unclear. Consequently, the necessity of the proposed decoding mechanism is not yet fully demonstrated.

Questions for authors
  1. What is the key methodological advantage of HyperReCo over closely related methods such as HGRAG and GFM-RAG? Could the authors compare HyperGNN against a standard query-conditioned document–entity bipartite GNN with matched supervision, input features, and model capacity? This would help isolate the contribution of the proposed architecture.

  2. Figure 3 examines message passing and query conditioning, but does not isolate the contributions of entity-level supervision and KL-based semantic distillation. Could the authors provide ablations of these objectives under the same HyperGNN architecture? This would help clarify which components drive the retrieval improvements.

  3. How faithfully do the gradient-based bridge scores in Eq. (6) capture specific cross-document dependencies? Could the authors validate this by masking high-attribution connections in the frozen HyperGNN and comparing their effects on target-document relevance with low-attribution or random connections?

Overall, I lean toward weak acceptance given the consistent empirical improvements. My main reservations concern methodological novelty and the independent contribution of GGHD. Convincing responses to these questions would strengthen my confidence in supporting acceptance.

6 comments
AuthorsOpenReview ↗

Response to Reviewer QK1C [Part 1/4]

We sincerely thank you for your positive and detailed review of our paper. We address the reviewer’s concerns point-by-point below.

Methodological contribution beyond HGRAG and GFM-RAG (Weakness 1 & Question 1)

We thank the reviewer for raising this concern. We clarify that our novelty lies in a novel framework that introduces a query-conditioned HyperGNN to learn higher-order evidence associations and translates learned structural relevance into explicit hyper-paths for LLM reasoning through gradient-guided hyper-path decoding. Key advantages over closely related methods are summarized as follows.

  • HGRAG. We adopt HGRAG's hypergraph and extraction protocol. HGRAG propagates relevance through predefined diffusion and leaves evidence connections implicit, whereas HyperReCo learns query-dependent propagation under joint supervision and decodes explicit connections among the retrieved evidence.
  • GFM-RAG. GFM-RAG learns message passing over a triple-based knowledge graph, which requires explicit relation extraction (48M vs. 3.3M LLM tokens per 10K documents, Table 4). HyperReCo operates directly on document-induced hyperedges and scores documents as evidence units.
  • HyperRAG. HyperRAG scores the plausibility of candidates during traversal over n-ary facts. In contrast, HyperReCo directly constructs document-induced hyperedges from extracted entities without explicit n-ary relation extraction, learns query-conditioned document relevance through hypergraph message passing, and derives explicit evidence paths.

Here, we provide additional theoretical analyses to further substantiate the advantages of our approach from two perspectives:

  1. Why document-induced hypergraphs? Document-induced hypergraphs preserve higher-order entity associations within documents and evidence connections across documents, providing richer structural information for document retrieval than pairwise KG representations that discard document membership. We theoretically demonstrate that distinct evidence configurations can yield identical pairwise KG representations, making them indistinguishable to retrievers operating solely on these representations (Proposition 1). This motivates using document-induced hypergraphs to provide more informative evidence structures for GNN-based retrieval.
  2. Why learned HyperGNN? Existing hypergraph-based methods (e.g., HGRAG) rely on predefined diffusion to propagate relevance over higher-order evidence associations. However, fixed propagation may fail to fully exploit these associations, as it can collapse distinct entity-level relevance distributions into identical document scores. Since message passing is central to HyperGNNs, we abstract HyperReCo's message-passing mechanism into a generalized hypergraph message-passing (HGMP) framework for theoretical analysis. We prove that HGMP can reproduce fixed diffusion while capturing nonlinear relevance interactions that fixed diffusion cannot distinguish (Theorem 1). This motivates our learned, query-conditioned HyperGNN to adaptively capture informative high-order interactions for multi-hop evidence retrieval.

The core proofs are presented below, and we have added detailed theoretical analysis and proofs in Appendix D.2.

Let HH be the entity–document incidence matrix, DVD_V the entity-degree matrix, and A(H)=HH⊤−DVA(\mathcal H)=HH^\top-D_V the pairwise projection. Define the document-level query coverage as C(H,Q)=∑d1[Q⊆ed]C(\mathcal H,Q)=\sum_d\mathbf1[Q\subseteq e_d]. There exist two hypergraphs with identical pairwise projections and query-dependent entity features but different query coverage.

Proposition 1 (Representational Advantage of Hypergraphs for GNN-Based Evidence Retrieval). Let C(H,Q)C(\mathcal H,Q) denote the document-level query coverage function. There exist two document-induced hypergraphs HA\mathcal H_A and HB\mathcal H_B with identical pairwise projections and query-dependent entity features, such that A(HA)=A(HB),C(HA,Q)≠C(HB,Q).A(\mathcal H_A)=A(\mathcal H_B),\qquad C(\mathcal H_A,Q)\neq C(\mathcal H_B,Q).

Proof. Consider two hypergraphs on V=1,…,6\mathcal V=\\{1,\ldots,6\\}: EA=1,2,3,1,4,5,2,4,6,3,5,6\mathcal{E}_A = \\{\\{1,2,3\\},\\{1,4,5\\},\\{2,4,6\\},\\{3,5,6\\}\\}

EB=1,2,4,1,3,5,2,3,6,4,5,6\mathcal{E}_B = \\{\\{1,2,4\\},\\{1,3,5\\},\\{2,3,6\\},\\{4,5,6\\}\\}

In both hypergraphs, every entity pair co-occurs exactly once, except 1,6,2,5,3,4\\{1,6\\},\\{2,5\\},\\{3,4\\}, which never co-occur. Therefore, A(HA)=A(HB).A(\mathcal H_A)=A(\mathcal H_B).

For Q=1,2,3Q=\\{1,2,3\\}, assign identical query-dependent entity features xv=1[v∈Q]x_v=\mathbf1[v\in Q]. Any deterministic KG-GNN gg operating solely on the pairwise projection and these features must satisfy

g(A(HA),X)=g(A(HB),X).g(A(\mathcal H_A),X)=g(A(\mathcal H_B),X).

However, HA\mathcal H_A contains a document covering all query entities, whereas HB\mathcal H_B does not:

C(HA,Q)=1≠0=C(HB,Q).\boxed{C(\mathcal H_A,Q)=1\neq0=C(\mathcal H_B,Q).} □\square

AuthorsOpenReview ↗

Response to Reviewer QK1C [Part 2/4]

Proposition 1 concerns what the hypergraph preserves. We now further examine how a retrieval mechanism uses query-relevance signals on hypergraph. Existing hypergraph-based methods (e.g., HGRAG) rely on predefined diffusion to propagate relevance over these higher-order evidence associations. However, fixed propagation may fail to fully exploit these associations, as it can collapse distinct entity-level relevance distributions into identical document scores.

Theorem 1 (Coverage and strict separation from fixed diffusion). For every valid HH, every pp, and every T≥0T\ge 0,

FdiffT(H,p)⊆FHGMPT(H,p).\mathcal F_{\mathrm{diff}}^{T}(H,p)\subseteq\mathcal F_{\mathrm{HGMP}}^{T}(H,p).

If, in addition, HH contains two distinct entities with identical hyperedge memberships, the inclusion is strict:

FdiffT(H,p)⊊FHGMPT(H,p).\boxed{\mathcal F_{\mathrm{diff}}^{T}(H,p)\subsetneq\mathcal F_{\mathrm{HGMP}}^{T}(H,p).}

Proof. Step 1: coverage. Choose linear HGMP messages and readout such that

u^{(t)}&=W_pD_E^{-1}H^\top D_V^{-1/2}h^{(t)},\ h^{(t+1)}&=D_V^{-1/2}Hu^{(t)},\ \widehat y^{(t)}&=W_pH^\top h^{(t)}. \end{aligned}$$ Substituting the first equation into the second yields $$h^{(t+1)}=S_qh^{(t)}.$$ With $h^{(0)}=x$, induction gives $$h^{(t)}=S_q^tx \quad\Longrightarrow\quad \boxed{\widehat y^{(t)}=W_pH^\top S_q^tx=y^{(t)}.}$$ Therefore, fixed hypergraph diffusion is exactly recoverable as a special case of HGMP. *Step 2: Strict separation.* Suppose two distinct nonisolated entities $v_1,v_2$ have identical hyperedge memberships. Let $z=\mathbf e_{v_1}-\mathbf e_{v_2}$. Since their incidence rows and degrees coincide, $$H^\top D_V^{-1/2}z=0 \quad\Longrightarrow\quad S_qz=0,\qquad W_pH^\top z=0.$$ For any $x'=x+\alpha z$, the diffusion scores satisfy $$\begin{aligned} y^{(t)}(x')-y^{(t)}(x)&=\alpha W_pH^\top S_q^tz\\ &=0,\qquad \forall t\geq0. \end{aligned}$$ Therefore, every deterministic post-processing of the diffusion scores takes the same value on $x$ and $x'$. Now consider two relevance vectors $x_c,x_s$, identical elsewhere, with $$(x_{c,v_1},x_{c,v_2})=(0,1), \qquad (x_{s,v_1},x_{s,v_2})=(\frac12,\frac12).$$ Since $x_s=x_c+\frac12z$, their diffusion scores are identical at every depth. For any hyperedge $e_d$ containing these entities, consider a GELU-based nonlinear HGMP readout with uniform attention: $$F_d(x)=\frac1{\delta_d}\sum_{v\in e_d} \operatorname{GELU}(x_v-\frac12),$$ where $\delta_d=|e_d|$. For any hyperedge containing $v_1,v_2$, using $\operatorname{GELU}(s)=s\Phi(s)$, we obtain $$\begin{aligned} F_d(x_c)-F_d(x_s)&=\frac{1}{\delta_d}\left[\operatorname{GELU}\left(\frac12\right)+\operatorname{GELU}\left(-\frac12\right) -2\operatorname{GELU}(0)\right]\\ &=\frac{1}{\delta_d} \left[\Phi\left(\frac12\right)-\frac12\right]>0. \end{aligned}$$ Therefore, we can derive $$\boxed{ y^{(t)}(x_c)=y^{(t)}(x_s),\ \forall t\geq0, \qquad F_d(x_c)\neq F_d(x_s). }$$ $\square$ ## Isolating the contributions of entity-level supervision and KL-based semantic distillation (Question 2) Thank you for this helpful suggestion. To isolate these objectives, we retrained the same HyperGNN architecture after removing either the entity-supervision loss (*w/o Entity*) or the KL-based semantic distillation loss (*w/o KL*), with all other settings fixed. Following Figure 3, we report R@5. From the results, **both objectives contribute to every dataset.** Removing entity supervision lowers mean R@5 by 2.1 points, with the largest effect on 2Wiki (2.9 points), and removing KL distillation lowers it by 1.4 points, with the largest effect on MuSiQue (2.9 points). *Ablation of the training objectives* | Variant | HotpotQA | MuSiQue | 2Wiki | Mean R@5 | | :---- | :---- | :---- | :---- | :---- | | HyperReCo | 96.8 | 74.5 | 97.8 | **89.7** | | w/o Entity | 95.2 | 72.6 | 94.9 | 87.6 | | w/o KL | 96.3 | 71.6 | 97.0 | 88.3 | **We have added these results in Section 5.3.**
AuthorsOpenReview ↗

Response to Reviewer QK1C [Part 3/4]

Comparison with bipartite GNNs (Question 1)

We thank the reviewer for bringing up the connection with a standard query-conditioned document–entity bipartite GNN. We want to clarify that our HyperGNN is a specialized, type-aware bipartite GNN with direction-specific, query-conditioned message passing (detailed proof in Appendix C.8.1). This design explicitly models entity–document membership roles and captures complementary evidence interactions essential for multi-hop retrieval. It offers more specialized representations than homogeneous bipartite GNNs, while remaining expressively equivalent in principle to sufficiently expressive heterogeneous bipartite GNNs. We further provide a theoretical analysis to clarify these expressive relationships.

(1) HyperGNN is a heterogeneous bipartite GNN. If document and entity nodes have type-specific message-passing and aggregation functions, a heterogeneous bipartite GNN can emulate our HyperGNN, which follows the two-stage hypergraph message-passing formulation in AllSet (Chien et al., 2022) and the incidence-graph view of UniGNN (Huang and Yang, 2021). We have added detailed theoretical analysis in Appendix C.8.1

(2) Advantages over Homogeneous Bipartite GNN. To evaluate the practical benefit of explicitly distinguishing the two propagation directions, we compared HyperGNN with a standard query-conditioned bipartite GNN (Homogeneous Bipartite GNN), which treats documents and entities as a single node type with shared message-passing and aggregation functions. Both models use the same input features, supervision, number of layers, and hidden dimension. From the results, HyperGNN improves average R@2 by 4.6 points and average R@5 by 3.0 points over the matched baseline. This is empirical evidence that explicitly distinguishing document and entity roles through direction-specific message passing benefits multi-hop evidence retrieval.

Retrieval recall of the HyperGNN and a matched homogeneous bipartite GNN

Avg. R@2Avg. R@5
Homogeneous BiGNN70.686.7
HyperReCo75.289.7

We have added the experimental results and detailed discussion in the Appendix C.8.2 of the revised manuscript.

Independent contribution of GGHD (Weakness 2)

We thank the reviewer for this important concern. We agree that GGHD yields only modest EM improvements over S-forest and G-paths. However, the primary objective of GGHD is to construct high-quality, compact, and useful hyper-paths. Since EM measures final answer correctness, it cannot directly capture the quality or usefulness of the generated evidence paths.

To evaluate this objective more directly, we employ LLM-as-a-judge to assess path-level helpfulness. GGHD achieves the highest helpful-path ratio (59.2%) while generating the fewest paths (1.68 per question), compared with 50.6% helpfulness and 2.11 paths for S-forest. These results suggest that GGHD produces more compact and useful evidence connections, providing complementary evidence beyond answer-level EM. To further validate the reliability of LLM-based evaluation, two PhD researchers independently annotated the decoded paths from 300 QA samples. Their helpful-path ratios were 66.4% and 68.6%, respectively, with over 82% agreement with the LLM judge (83.5% on average). These results demonstrate encouraging consistency between human and LLM judgments, further supporting the usefulness of GGHD-generated evidence paths.

Table 11. LLM-judged path helpfulness and number of paths

MethodHelpful-path ratio (%) ↑Mean paths per question ↓
GGHD59.21.68
w/o Q-PCSF28.45.39
G-paths43.12.97
S-forest50.62.11

GGHD adopts gradient-based attribution to identify evidence connections explicitly. Specifically, gradient attribution estimates the sensitivity of predicted relevance to candidate bridges, while Q-PCSF globally organizes these connections into compact hyper-paths. To validate this design, we conduct frozen-model message masking (detailed in our response to Question 3). Masking high-attribution bridges causes substantially larger target-logit drops than masking random or low-attribution bridges (0.154 vs. 0.067 and 0.017 on average), supporting the effectiveness of gradient attribution in identifying prediction-sensitive connections.

We have added the detailed discussion in Appendices C.4, C.5 of the revised manuscript.

Reviewer_vZyAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a hypergraph-based retrieval framework for improving LLM multi-hop question answering. The central idea is to represent documents as hyperedges and entities as nodes, thereby capturing higher-order entity associations within documents and connections across documents through shared entities. Experiments demonstrate competitive retrieval and QA performance.

Critical strengths
  1. The research problem is important and well-motivated. The paper correctly identifies that retrieving relevant documents does not necessarily give an LLM enough guidance to combine their evidence. Explicitly representing connections among retrieved documents is a meaningful direction for improving multi-hop reasoning.

  2. The proposed framework combines query-conditioned hypergraph message passing with explicit hyper-path construction. This design provides a systematic connection between learned structural relevance and the evidence organization presented to the LLM.

  3. The indexing cost is relatively low. The reported indexing cost is considerably lower than several graph-based baselines, making the approach attractive for retrieval applications.

Critical weaknesses
  1. The methodological novelty of the HyperGNN retriever seems limited. I don't see a need to consider hyperedges over standard heterogeneous or bipartite GNNs. In addition, the query-conditioned message passing, attention-based aggregation, and joint document/entity relevance supervision are all established techniques. The authors should clarify which architectural components constitute genuinely new contributions relative to existing learning-based graph retrievers.

  2. The technical part presents what was done without much explanation of why it should be done this way. For example, Equation 6 clearly shows how the directional bridge attribution is calculated, but I don't understand why it should be calculated this way. Those derivatives measure the local sensitivity of the prediction to individual messages. They do not necessarily measure the contribution of the complete two-step information flow from the source document to the target document.

  3. The experiments use GPT-4o-mini as the downstream reasoning model. It is unclear whether and how the performance depends on specific LLMs or multiple LLMs as a group.

Questions for authors
  1. What specific higher-order interactions can the proposed HyperGNN represent that a query-conditioned bipartite document–entity GNN cannot capture under comparable model capacity?

  2. What is the rationale for the design of Equation 6?

  3. To what extent does the improvement from hyper-path guidance depend on the reasoning capability of the downstream LLM?

7 comments
AuthorsOpenReview ↗

Response to Reviewer vZyA [Part 1/3]

We sincerely thank you for your detailed and insightful review of our paper. Below, we address your concerns point-by-point.

Methodological novelty of the HyperGNN retriever (Weakness 1)

We thank the reviewer for raising this concern. We want to clarify that our novelty lies in a novel framework that introduces a query-conditioned HyperGNN to learn higher-order evidence associations and translates learned structural relevance into explicit hyper-paths for LLM reasoning through gradient-guided hyper-path decoding. Compared with existing graph retrievers, HyperReCo preserves document-level evidence associations without explicit relation extraction and adaptively models query-dependent entity–document interactions to support multi-hop evidence retrieval. Here, we provide additional theoretical analyses to further substantiate the advantages of our approach from two perspectives:

  1. Why document-induced hypergraphs? Document-induced hypergraphs preserve higher-order entity associations within documents and evidence connections across documents, providing richer structural information for document retrieval than pairwise KG representations that discard document membership. We theoretically demonstrate that distinct evidence configurations can yield identical pairwise KG representations, making them indistinguishable to retrievers operating solely on these representations (Proposition 1). This motivates using document-induced hypergraphs to provide more informative evidence structures for GNN-based retrieval.
  2. Why learned HyperGNN? Existing hypergraph-based methods (e.g., HGRAG) rely on predefined diffusion to propagate relevance over higher-order evidence associations. However, fixed propagation may fail to fully exploit these associations, as it can collapse distinct entity-level relevance distributions into identical document scores. Since message passing is central to HyperGNNs, we abstract HyperReCo's message-passing mechanism into a generalized hypergraph message-passing (HGMP) framework for theoretical analysis. We prove that HGMP can reproduce fixed diffusion while capturing nonlinear relevance interactions that fixed diffusion cannot distinguish (Theorem 1). This motivates our learned, query-conditioned HyperGNN to adaptively capture informative high-order interactions for multi-hop evidence retrieval.

The core proofs are presented below, and we have added detailed theoretical analysis and proofs in Appendix D.2.

Let HH be the entity–document incidence matrix, DVD_V the entity-degree matrix, and A(H)=HH⊤−DVA(\mathcal H)=HH^\top-D_V the pairwise projection. Define the document-level query coverage as C(H,Q)=∑d1[Q⊆ed]C(\mathcal H,Q)=\sum_d\mathbf1[Q\subseteq e_d]. There exist two hypergraphs with identical pairwise projections and query-dependent entity features but different query coverage.

Proposition 1 (Representational Advantage of Hypergraphs for GNN-Based Evidence Retrieval). Let C(H,Q)C(\mathcal H,Q) denote the document-level query coverage function. There exist two document-induced hypergraphs HA\mathcal H_A and HB\mathcal H_B with identical pairwise projections and query-dependent entity features, such that A(HA)=A(HB),C(HA,Q)≠C(HB,Q).A(\mathcal H_A)=A(\mathcal H_B),\qquad C(\mathcal H_A,Q)\neq C(\mathcal H_B,Q). Proof. Consider two hypergraphs on V=1,…,6\mathcal V=\\{1,\ldots,6\\}: EA=1,2,3,1,4,5,2,4,6,3,5,6\mathcal{E}_A = \\{\\{1,2,3\\},\\{1,4,5\\},\\{2,4,6\\},\\{3,5,6\\}\\} EB=1,2,4,1,3,5,2,3,6,4,5,6\mathcal{E}_B = \\{\\{1,2,4\\},\\{1,3,5\\},\\{2,3,6\\},\\{4,5,6\\}\\} In both hypergraphs, every entity pair co-occurs exactly once, except 1,6,2,5,3,4\\{1,6\\},\\{2,5\\},\\{3,4\\}, which never co-occur. Therefore, A(HA)=A(HB).A(\mathcal H_A)=A(\mathcal H_B). For Q=1,2,3Q=\\{1,2,3\\}, assign identical query-dependent entity features xv=1[v∈Q]x_v=\mathbf1[v\in Q]. Any deterministic KG-GNN gg operating solely on the pairwise projection and these features must satisfy g(A(HA),X)=g(A(HB),X).g(A(\mathcal H_A),X)=g(A(\mathcal H_B),X). However, HA\mathcal H_A contains a document covering all query entities, whereas HB\mathcal H_B does not: C(HA,Q)=1≠0=C(HB,Q).\boxed{C(\mathcal H_A,Q)=1\neq0=C(\mathcal H_B,Q).} □\square Proposition 1 concerns what the hypergraph preserves. We now further examine how a retrieval mechanism uses query-relevance signals on hypergraph. Existing hypergraph-based methods (e.g., HGRAG) rely on predefined diffusion to propagate relevance over these higher-order evidence associations. However, fixed propagation may fail to fully exploit these associations, as it can collapse distinct entity-level relevance distributions into identical document scores.

AuthorsOpenReview ↗

Response to Reviewer vZyA [Part 2/3]

Theorem 1 (Coverage and strict separation from fixed diffusion). For every valid HH, every pp, and every T≥0T\ge 0,

FdiffT(H,p)⊆FHGMPT(H,p).\mathcal F_{\mathrm{diff}}^{T}(H,p)\subseteq\mathcal F_{\mathrm{HGMP}}^{T}(H,p).

If, in addition, HH contains two distinct entities with identical hyperedge memberships, the inclusion is strict:

FdiffT(H,p)⊊FHGMPT(H,p).\boxed{\mathcal F_{\mathrm{diff}}^{T}(H,p)\subsetneq\mathcal F_{\mathrm{HGMP}}^{T}(H,p).}

Proof. Step 1: coverage. Choose linear HGMP messages and readout such that

u^{(t)}&=W_pD_E^{-1}H^\top D_V^{-1/2}h^{(t)},\ h^{(t+1)}&=D_V^{-1/2}Hu^{(t)},\ \widehat y^{(t)}&=W_pH^\top h^{(t)}. \end{aligned}$$ Substituting the first equation into the second yields $$h^{(t+1)}=S_qh^{(t)}.$$ With $h^{(0)}=x$, induction gives $$h^{(t)}=S_q^tx \quad\Longrightarrow\quad \boxed{\widehat y^{(t)}=W_pH^\top S_q^tx=y^{(t)}.}$$ Therefore, fixed hypergraph diffusion is exactly recoverable as a special case of HGMP. *Step 2: Strict separation.* Suppose two distinct nonisolated entities $v_1,v_2$ have identical hyperedge memberships. Let $z=\mathbf e_{v_1}-\mathbf e_{v_2}$. Since their incidence rows and degrees coincide, $$H^\top D_V^{-1/2}z=0 \quad\Longrightarrow\quad S_qz=0,\qquad W_pH^\top z=0.$$ For any $x'=x+\alpha z$, the diffusion scores satisfy $$\begin{aligned} y^{(t)}(x')-y^{(t)}(x)&=\alpha W_pH^\top S_q^tz\\ &=0,\qquad \forall t\geq0. \end{aligned}$$ Therefore, every deterministic post-processing of the diffusion scores takes the same value on $x$ and $x'$. Now consider two relevance vectors $x_c,x_s$, identical elsewhere, with $$(x_{c,v_1},x_{c,v_2})=(0,1), \qquad (x_{s,v_1},x_{s,v_2})=(\frac12,\frac12).$$ Since $x_s=x_c+\frac12z$, their diffusion scores are identical at every depth. For any hyperedge $e_d$ containing these entities, consider a GELU-based nonlinear HGMP readout with uniform attention: $$F_d(x)=\frac1{\delta_d}\sum_{v\in e_d} \operatorname{GELU}(x_v-\frac12),$$ where $\delta_d=|e_d|$. For any hyperedge containing $v_1,v_2$, using $\operatorname{GELU}(s)=s\Phi(s)$, we obtain $$\begin{aligned} F_d(x_c)-F_d(x_s)&=\frac{1}{\delta_d}\left[\operatorname{GELU}\left(\frac12\right)+\operatorname{GELU}\left(-\frac12\right) -2\operatorname{GELU}(0)\right]\\ &=\frac{1}{\delta_d} \left[\Phi\left(\frac12\right)-\frac12\right]>0. \end{aligned}$$ Therefore, we can derive $$\boxed{ y^{(t)}(x_c)=y^{(t)}(x_s),\ \forall t\geq0, \qquad F_d(x_c)\neq F_d(x_s). }$$ $\square$ ## Comparison with bipartite GNNs (Question 1) We thank the reviewer for bringing up the connection with bipartite GNNs. We want to clarify that our HyperGNN is **a specialized, type-aware bipartite GNN** with direction-specific, query-conditioned message passing (detailed proof in **Appendix C.8.1**). This design explicitly models entity–document membership roles and captures complementary evidence interactions essential for multi-hop retrieval. It offers more specialized representations than homogeneous bipartite GNNs, while remaining expressively equivalent in principle to sufficiently expressive heterogeneous bipartite GNNs. **We further provide a theoretical analysis to clarify these expressive relationships.** (1) **HyperGNN is a heterogeneous bipartite GNN.** If document and entity nodes have type-specific message-passing and aggregation functions, **a heterogeneous bipartite GNN can emulate our HyperGNN**, which follows the two-stage hypergraph message-passing formulation in AllSet (Chien et al., 2022) and the incidence-graph view of UniGNN (Huang and Yang, 2021). **We have added detailed theoretical analysis in Appendix C.8.1** (2) **Advantages over Homogeneous Bipartite GNN.** To evaluate the practical benefit of explicitly distinguishing the two propagation directions, we compared HyperGNN with a standard query-conditioned bipartite GNN (Homogeneous Bipartite GNN), which treats documents and entities as a single node type with shared message-passing and aggregation functions. Both models use the same input features, supervision, number of layers, and hidden dimension. From the results, HyperGNN improves average R@2 by 4.6 points and average R@5 by 3.0 points over the matched baseline. **This is empirical evidence that explicitly distinguishing document and entity roles through direction-specific message passing benefits multi-hop evidence retrieval.** *Retrieval recall of the HyperGNN and a matched homogeneous bipartite GNN* | | Avg. R@2 | Avg. R@5 | | :---- | :---- | :---- | | Homogeneous BiGNN | 70.6 | 86.7 | | HyperReCo | **75.2** | **89.7** | **We have added the experimental results and detailed discussion in the Appendix C.8.2 of the revised manuscript.**
AuthorsOpenReview ↗

Response to Reviewer vZyA [Part 3/3]

Rationale for Eq. (6) (Weakness 2 & Question 2)

Thank you for this thoughtful comment. We agree that the manuscript should better motivate this design, and we have clarified it below.

Rationale for Eq. (6). For an entity-mediated connection between two documents, two directional relationships determine how evidence propagates from one document to the other: (1) the contribution of the source document to the shared entity, and (2) the contribution of the shared entity to the target document. Our query-conditioned HyperGNN captures these two relationships through its two-stage message passing (Eqs. 3–4). Building on the trained model, Eq. (6) uses gradient-based attribution to quantify the influence of both directional messages on the relevance prediction of the target document. Importantly, the gradient with respect to the source-to-entity message already captures its downstream influence via the entity-to-target propagation through the chain rule. Averaging the two directional attributions therefore gives a first-order sensitivity measure that accounts for both stages of the entity-mediated connection.

What Eq. (6) measures. Following gradient-based path interpretation in GNNs (NBFNet, GFM-RAG), Eq. (6) explains the trained HyperGNN: it identifies which entity-mediated connections the trained model relies on when predicting the relevance of target document djd_j for query qq. Specifically, a connection edi→v→edje_{d_i} \to v \to e_{d_j} consists of two consecutive messages, controlled by the gates $ \gamma_1 = \gamma^{(\ell)}_{e_{d_i}\to v} $, and $ \gamma_2 = \gamma^{(\ell)}_{v\to e_{d_j}} $. Then, eq. (6) scores both steps against the same target logit sdjs_{d_j}.

  • Step 1, source document to bridge entity (edi→ve_{d_i} \to v). ∂sdj/∂γ1\partial s_{d_j}/\partial\gamma_1 measures how much the target prediction depends on what did_i writes into vv. This is not a one-step quantity: the message updates hv(ℓ+1)h_v^{(\ell+1)}, which the second message then uses (Eqs. 3–4), so by the chain rule the derivative already carries the effect of the source message through v→edjv \to e_{d_j} to the target logit.
  • Step 2, bridge entity to target document (v→edjv \to e_{d_j}). ∂sdj/∂γ2\partial s_{d_j}/\partial\gamma_2 measures how much the target prediction depends on what vv delivers into djd_j, which includes the content vv has just received from did_i.

Empirical support. Replacing gradient-derived costs with semantic-similarity costs (S-forest) reduces the LLM-judged helpful-path ratio from 59.2% to 50.6% despite producing more paths (2.11 vs. 1.68 per question), and slightly lowers mean EM from 57.6 to 57.0 (Tables 10 and 11). In addition, on 100 questions per dataset, paths built from high-attribution connections outperform those built from random or low-attribution connections in both mean EM (53.3 vs. 50.3 and 48.7) and judged helpfulness (36.0% vs. 32.4% and 29.2%). These measures reflect downstream usefulness rather than faithfulness to the model, so we regard them as supporting evidence.

We have added this detailed discussion to Section 4.2 and Appendix D.1.

Robustness to different reader LLMs (Weakness 3 & Question 3).

Thank you for this suggestion. We replaced GPT-4o-mini with two additional reader LLMs, Llama-3.3-70B-Instruct and Qwen3.8-Flash, and kept the rest of the pipeline unchanged. Results are reported as EM / F1. From the results, HyperReCo outperforms EHRAG and HGRAG in both EM and F1 on all three datasets under both additional LLMs. Its average EM advantage over HGRAG is 3.6 points with Llama-3.3-70B-Instruct and 9.0 points with Qwen3.8-Flash, compared with 4.2 points with GPT-4o-mini, indicating that the improvement is not specific to one reader. Meanwhile, hyper-paths improve EM in all nine LLM–dataset combinations, which demonstrates that hyper-path guidance remains beneficial across all three readers.

QA performance (EM / F1) with two additional reader LLMs.

MethodHotpotQAMuSiQue2WikiAvg.
EMF1EMF1EMF1EMF1
Llama-3.3-70B-Instruct
EHRAG60.172.935.745.465.671.653.863.3
HGRAG61.274.037.547.670.276.856.366.1
HyperReCo64.177.139.549.876.182.759.969.9
w/o hyper-paths63.876.836.446.975.281.558.568.4
Qwen3.8-Flash
EHRAG57.669.132.740.260.364.750.258.0
HGRAG58.970.833.039.562.667.851.559.3
HyperReCo65.578.539.647.476.481.660.569.2
w/o hyper-paths61.473.336.343.576.380.958.065.9

We have added these results in the Appendix C.1 of the revised manuscript.

Reviewer_Bfd6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents HyperReCo, which retrieves complementary documents and makes their connections explicit for LLM multi-hop reasoning. It represents each document as a hyperedge over extracted entities, trains a query-conditioned HyperGNN with document and entity supervision, and decodes gradient-guided hyper-paths among retrieved documents. The path decoder attributes entity-mediated messages to document relevance and selects a forest through a query-conditioned prize-collecting Steiner objective. Experiments on three multi-hop QA datasets and three domain-specific GraphRAG benchmarks show strong retrieval and QA performance; ablations indicate benefits from query-conditioned message passing, gradient attribution, and forest decoding.

Critical strengths
  • The method addresses both evidence retrieval and evidence routing. HyperReCo does not stop at ranking documents: gradient-guided decoding converts learned entity-mediated interactions into explicit document paths supplied to the LLM. This directly targets a common failure in graph RAG, where relevant documents are retrieved but their composition is left implicit.

  • The evaluation covers retrieval, downstream reasoning, robustness, and efficiency. The paper reports six benchmarks, strong dense and graph-based baselines, four embedding backbones, component ablations, alternative path decoders, latency and indexing cost, an LLM-judged path analysis, and a qualitative case. The additional path analyses are particularly relevant to the central claim.

Critical weaknesses
  • Hypergraph-construction quality is not evaluated. The entire retrieval and path-decoding pipeline depends on LLM-based entity extraction and canonicalization: extraction errors, missed aliases, or over-merged entities directly change hyperedges, shared entities, and the learned message-passing structure. The paper follows HGRAG’s extraction protocol but reports no entity-extraction precision/recall, canonicalization-error analysis, or sensitivity to noisy entities.

  • The scalability and reliability of decoded paths need stronger support. The exact forest solver enumerates candidate forests over only five retrieved documents, and the paper does not analyze larger evidence sets. Paths are based on shared entities and gradient sensitivity rather than validated relational semantics; even the LLM-judged evaluation labels only 59.2% of GGHD paths as helpful, and that evaluation has no human-agreement check.

Questions for authors
  • How does exact forest decoding scale beyond five retrieved documents? Is there an approximate algorithm with similar accuracy and lower combinatorial cost?

  • What is human agreement with the LLM-judged path-helpfulness labels, and how often do spurious shared-entity paths change the final answer?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Bfd6 [Part 1/2]

We sincerely thank you for your positive and detailed review. Below, we address each concern point-by-point.

Hypergraph-construction quality (Weakness 1)

We thank the reviewer for this valuable suggestion. We agree that entity extraction and canonicalization shape the hyperedges and shared entities on which HyperReCo operates. However, similar to prior work such as GFM-RAG and HGRAG, hypergraph-construction quality is not the primary focus of our work. Our contribution lies in learning and decoding structural relevance over the constructed hypergraph. We adopt HGRAG's extraction protocol and re-run HGRAG with the same LLM for graph construction. We further conduct experiments to quantify the quality of the constructed hypergraphs.

Extraction quality. Due to the lack of ground truth, we therefore adopt LLM-as-a-judge (e.g., GPT-4o) to quantify the extraction quality. Specifically, we randomly sampled 100 documents from each of HotpotQA, MuSiQue, and 2Wiki, and used GPT-4o to assess the extracted entities against the source text on three metrics: Entity Precision (the proportion of extracted entities judged valid), Entity Coverage (the proportion of important entities in the document that were extracted, serving as a proxy for recall), and Unsupported Entity Rate (the proportion of extracted entities not supported by the document).

The results indicate that the extracted entities are generally reliable, with 91.0% precision, 92.2% coverage, and a 7.4% unsupported rate overall. The automatically constructed hypergraphs are therefore reliable enough to support effective multi-hop evidence retrieval, as reflected by the results in Table 2.

LLM-judged hypergraph quality.

Dataset#DocsPrecision (%)Coverage (%)Unsupported (%)
HotpotQA10092.895.46.0
MuSiQue10090.992.87.7
2Wiki10089.488.48.5
Overall30091.092.27.4

We have added these results in Appendix C.7 of the revised manuscript.

Scalability of exact forest decoding beyond five retrieved documents (Weakness 2 & Question 1)

Thank you for this constructive comment. We evaluated HyperReCo with K∈{5,8,10,15,20}K\in\{5, 8, 10, 15, 20\} retrieved documents on 100 randomly sampled questions per dataset, using the exact solver throughout. From the results, increasing KK from 5 to 8 improves mean EM from 57.0 to 59.0 and mean F1 from 66.8 to 69.5, indicating that HyperReCo can benefit from a moderately larger evidence set. Mean EM is lower at K≥15K \ge 15, which is consistent with additional non-supporting documents adding noise for the reader.

QA performance and mean exact-solver time per question.

KHotpotQAMuSiQue2WikiMeanMean solver time (s), H / M / 2W
EMF1EMF1EMF1EMF1
560.074.038.046.973.079.657.066.80.01 / 0.01 / 0.01
864.076.436.048.877.083.159.069.50.12 / 0.07 / 0.10
1059.072.937.048.673.081.556.367.70.19 / 0.09 / 0.10
1560.075.136.047.470.080.955.367.80.23 / 0.40 / 0.23
2055.070.438.050.470.081.454.367.40.31 / 0.59 / 0.22

Solver cost. Exact decoding is exponential in the worst case, but in the number of candidate connections rather than in KK itself. In practice, this graph is sparse because we retain only document pairs that share entities and have positive gradient attribution. As a result, the mean solver time remains below 0.6 s even at K=20K=20, indicating an acceptable computational overhead.

Approximate decoding. For denser graphs or larger KK, the exact solver can be replaced with polynomial-time prize-collecting solvers, such as the 2-approximation for prize-collecting Steiner forest [1], or the nearly-linear-time factor-2 variant [2].

[1] Ahmadi, Gholami, Hajiaghayi, Jabbarzade, and Mahdavi. 2-Approximation for Prize-Collecting Steiner Forest. Journal of the ACM, 2025. [2] Hegde, Indyk, and Schmidt. A Nearly-Linear Time Framework for Graph-Structured Sparsity. ICML, 2015.

We have added these results and discussion in Appendix C.6 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Bfd6 [Part 2/2]

Human validation and the effect of spurious paths (Weakness 2 & Question 2)

Thank you for this constructive comment. We agree that the reliability of decoded hyper-paths is crucial for downstream reasoning. To address this concern, we clarify the effectiveness of hyper-path guidance and conduct a human agreement study to validate the reliability of LLM-based helpfulness evaluation.

(W2) Reliability of decoded paths. Hyper-paths are designed to explicitly represent evidence connections. The prompt (Figure 5) instructs the LLM to use them only as optional navigation and to verify every fact in the documents, so relational semantics remain grounded in the source text. Notably, GGHD achieves the highest helpfulness ratio (59.2% vs. 28.4-50.6%) while generating the fewest paths (1.68 on average), demonstrating its effectiveness in identifying compact and useful evidence connections. Furthermore, removing hyper-paths reduces mean EM from 57.6 to 55.0 (Table 11), confirming their contribution to downstream reasoning.

(Q2) Human agreement. We invited two PhD researchers to independently label the GGHD paths decoded for the total 300 QA samples used in Table 11. From the results, both annotators agree with the LLM judge on more than 82% of the paths (83.5% on average), which indicates that the LLM-judged labels are a reasonable proxy for human judgment.

Human validation of the LLM-judged helpfulness labels for GGHD paths

AnnotatorAgreement with LLM judge (%)Helpful-path ratio (%)
A184.166.4
A282.968.6
Mean83.567.5

We have added these results to Appendix C.4.2 of the revised manuscript.

Effect of spurious paths. To further investigate the impact of path quality on downstream reasoning, we kept the retrieved documents (K=5K=5) and the reader fixed, and replaced the GGHD hyper-paths with spurious paths built from low-attribution or random shared-entity connections on the same 100 questions per dataset. We evaluate both downstream QA performance and path helpfulness using the LLM judge described in Appendix C.4.1.

From the results, replacing GGHD paths with low-attribution or random paths reduces mean EM by 8.3 and 6.7 points, respectively, accompanied by substantially lower helpful-path ratios (29.2% and 32.4% vs. 59.2%). These results indicate that potentially spurious shared-entity connections can negatively affect downstream reasoning, highlighting the importance of selectively decoding reliable evidence connections.

Final-answer quality and judged helpfulness when GGHD hyper-paths are replaced by attribution-controlled path sets.

Path setHotpotQA EM / F1MuSiQue EM / F12Wiki EM / F1Mean EM / F1Helpful-path ratio (%)
GGHD60.0 / 74.038.0 / 46.973.0 / 79.657.0 / 66.859.2
Random56.0 / 70.434.0 / 44.461.0 / 69.550.3 / 61.432.4
Low-attribution52.0 / 67.833.0 / 44.161.0 / 71.848.7 / 61.329.2

We have added these results to Appendix C.5 of the revised manuscript.

AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer Bfd6,

We sincerely thank you for your valuable feedback. We would greatly appreciate your feedback on whether our responses have adequately addressed your concerns. If you find that our responses have addressed your concerns, we would be grateful if you would consider updating your rating accordingly.

Thank you again for your time and consideration.

Best regards, All authors

Dear Reviewer Bfd6,

As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our response and let us know whether your concerns have been adequately addressed. If you find that our responses have addressed your concerns, we would be grateful if you would consider updating your rating accordingly.

Thank you again for your time and consideration.

Best regards, All authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their thorough and constructive reviews of our paper. We are grateful for the positive feedback and constructive suggestions provided, which have helped us improve the clarity and quality of our work. We have carefully revised the paper according to the comments, and the edits have been highlighted in BLUE. Below, we summarize the major revisions and clarify our contribution. Detailed point-by-point responses are provided separately for each reviewer.

  • Novelty and Contributions (Reviewers vZyA and QK1C). We want to clarify that our contribution is a novel framework that introduces a query-conditioned HyperGNN to learn higher-order evidence associations and translates learned structural relevance into explicit hyper-paths for LLM reasoning through gradient-guided hyper-path decoding. We further demonstrate the effectiveness of our approach through both theoretical analysis and extensive experiments. We have added detailed discussion in the general response and in our response to Reviewers vZyA and QK1C.
  • Comparison with bipartite GNNs (Reviewers vZyA and QK1C). We clarify that HyperGNN is a specialized bipartite GNN with type-aware, query-conditioned message passing, and further provide theoretical analysis and evaluate its empirical advantages through a matched bipartite-GNN baseline. We have added a detailed discussion in Appendix C.8 of the revised manuscript and in our response to Reviewers vZyA and QK1C.
  • Rationale and faithfulness of Eq. (6) (Reviewers vZyA and QK1C). We have clarified that Eq. (6) measures the query-specific importance of entity-mediated document connections through the gradient sensitivity of relevance predictions, and further validated its faithfulness through additional masking experiments. We have added the detailed discussion and experimental results in Section 4.2 and Appendices C.5, D.1 and in our response to Reviewers vZyA and QK1C.
  • Independent contribution of GGHD (Reviewer QK1C). We have conducted additional experiments to demonstrate that GGHD more effectively leverages learned gradient information to construct useful and compact hyper-paths, and added results in Appendices C.4, C.5.
  • Reliability of decoded hyper-paths (Reviewers Bfd6 and QK1C). We have conducted additional experiments to validate the reliability of LLM-based helpfulness evaluation, and added results in Appendices C.4.2, C.5.
  • Scalability of forest decoding (Reviewer Bfd6). We have evaluated forest decoding with up to K=20K=20 retrieved evidence sets, and added additional results and analysis in Table 15 and Appendix C.6.
  • Impact on different reader LLMs (Reviewer vZyA). We have extended our evaluation to two additional reader LLMs, Llama-3.3-70B-Instruct and Qwen3.8-Flash, alongside GPT-4o-mini, and added results in Table 8 and Appendix C.1.
  • Hypergraph-construction quality (Reviewer Bfd6). We have conducted additional experiments to assess the quality of the constructed hypergraphs, and added results in Table 16 and Appendix C.7.
  • Training objectives (Reviewer QK1C). We have conducted additional ablation studies to evaluate the contributions of entity-level supervision and KL-based semantic distillation, and added results in Figure 3 and Section 5.3.

Novelty Contributions

We thank the reviewers for their constructive comments regarding methodological novelty. We would like to clarify that our contribution is a novel framework that introduces a query-conditioned HyperGNN to learn higher-order evidence associations and translates learned structural relevance into explicit hyper-paths for LLM reasoning through gradient-guided hyper-path decoding. Our additional theoretical analysis (see our responses to Reviewers vZyA and QK1C, and Appendix D.2) further substantiates this design: document-induced hypergraphs preserve higher-order evidence associations essential for multi-hop retrieval, while our learned, query-conditioned HyperGNN can adaptively distinguish entity-level relevance interactions that previous methods cannot capture. Built upon a document-induced hypergraph, our framework comprises two tightly connected components.

(1) Query-conditioned HyperGNN retriever. Our HyperGNN learns query-dependent document-entity interactions through trainable message passing and joint relevance supervision. Theoretical analysis demonstrates that our learned, query-conditioned HyperGNN can distinguish entity-level evidence interactions that predefined linear diffusion may collapse.

(2) Gradient-Guided Hyper-Path Decoding (GGHD). Beyond relevance scores, GGHD attributes each document prediction to entity-mediated messages and organizes the attributed connections into explicit hyper-paths through query-conditioned prize-collecting Steiner forest (Q-PCSF) optimization.

We hope this clarifies our perspective on the paper's novelty.

AuthorsOpenReview ↗

Summary of Rebuttal for the Area Chair

Dear Area Chair,

We thank you for handling our submission and the reviewers for their constructive feedback. All three reviewers now recommend acceptance, with scores of 3, 3, and 3. Notably, following our detailed rebuttal and additional analyses, Reviewer vZyA acknowledged that our responses helped address concerns and raised the score from 2 to 3. As the discussion period comes to an end, we provide a brief summary of the reviewers' key comments and our corresponding responses, which we hope will facilitate your final decision.

Strengths recognized by the reviewers. The reviewers highlighted the following strengths of HyperReCo:

  • An important, well-motivated research problem. All three reviewers agree that retrieving the relevant documents does not guarantee that an LLM can connect their evidence. They agree that making these connections explicit is a meaningful direction for multi-hop reasoning (Bfd6, vZyA, QK1C).
  • Joint evidence retrieval and connection. HyperReCo learns and decodes structural relevance over the constructed hypergraph. The proposed framework combines query-conditioned hypergraph message passing with explicit evidence hyper-paths for LLM reasoning (Bfd6, vZyA, QK1C). On the same retrieved documents, explicit connections raise mean EM from 55.0 to 57.6 (QK1C).
  • Strong, consistent results and comprehensive evaluation. HyperReCo consistently improves both retrieval recall and QA accuracy, with substantial gains over HGRAG and robustness across embedding backbones (QK1C). The evaluation covers six benchmarks, strong dense and graph-based baselines, four embedding backbones, component ablations, alternative path decoders and an efficiency analysis (Bfd6).
  • Low indexing cost. HyperReCo’s indexing cost is considerably lower than that of several graph-based baselines, which makes it attractive for retrieval applications (vZyA).
  • Clear presentation. All three reviewers rated the presentation as clear and easy to follow (Presentation: 3).

Concerns addressed in our rebuttal. We responded to every concern point by point, with new theoretical analysis and additional experiments:

  • Methodological novelty (vZyA, QK1C). We clarified the contributions of the HyperGNN and GGHD. This is supported by a new theoretical analysis (Appendix D.2) and a comparison with a matched bipartite-GNN baseline (Appendix C.8).
  • Rationale and faithfulness of the gradient attribution (vZyA, QK1C). We explained the motivation for Eq. (6) (Section 4.2, Appendix D.1) and confirmed its faithfulness with gradient masking experiments (Appendix C.5).
  • GGHD’s independent contribution and reliability (Bfd6, QK1C). Additional path-quality evaluations and a human annotation study show that the decoded hyper-paths are useful and reliable (Appendices C.4, C.5).
  • Further validation. We evaluated hypergraph-construction quality and forest-decoding scalability (Bfd6), generalization across reader LLMs (vZyA), and the contribution of each training objective (QK1C). Refer to Tables 8, 15 and 16 and Figure 3.

We believe our responses and the corresponding revisions to the manuscript (highlighted in blue) resolve their main concerns, as reflected as updated positive assessments from all three reviewers. We would be grateful if you could take these responses and revisions into account in your final recommendation. Thank you for your time and consideration.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.