acceptodds
Under review as a conference paper at ICLR 2027

Corpus-Guided Dual-Path Propagation for Graph Retrieval-Augmented Generation

Abstract

Graph-based retrieval-augmented generation supports multi-hop retrieval by organizing corpus information into graphs. However, existing relation-free graph retrieval methods rely primarily on query–sentence similarity to search for evidence. This can exclude useful bridging evidence with low query similarity and activate incidental entities unrelated to the reasoning chain. In this paper, we propose a simple and effective approach called NexusRAG, which augments the relation-free Tri-Graph with a corpus-level entity neighborhood structure derived from joint entity co-occurrence and semantic similarity. NexusRAG employs this structure to guide two complementary propagation paths: neighborhood-constrained semantic propagation through sentences identifies the query-relevant entity frontier, while direct structural propagation between neighboring entities expands that frontier to structurally related entities. The propagated entity weights also inform neighborhood-aware passage initialization for Personalized PageRank. Experiments on three multi-hop QA benchmarks and a domain-specific subset of GraphRAG-Bench show that NexusRAG consistently outperforms existing approaches. On the GraphRAG-Bench subset, NexusRAG achieves the highest evidence recall in all question categories, exceeding baselines by 4.2-8.1 points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large languagde models, retrieval augmented generation, graph retrieval augmented generation, multi-hop retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1BVwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors identify a limitation of LinearRAG: its reliance on query–sentence similarity for entity exploration can overlook critical bridge evidence with low query relevance or activate irrelevant entities. To address this, they propose NexusRAG, which constructs a Corpus-level Neighbor Prior by combining entity co-occurrence and semantic similarity. Using Dual-path Propagation, NexusRAG suppresses irrelevant entity activation while discovering potentially overlooked bridge entities. It then retrieves supporting passages through Neighbor-aware Passage Initialization and Personalized PageRank (PPR), incorporating the importance of propagated entities. Without requiring LLM-based relation extraction, NexusRAG improves multi-hop QA accuracy and evidence retrieval recall over LinearRAG.

Critical strengths
  • Clear Problem Formulation and Motivation

The paper identifies two limitations of entity propagation in LinearRAG: query-gated cutoff and spurious activation. Reliance on query–sentence similarity can either overlook essential intermediate evidence or activate irrelevant entities, both of which are critical challenges in multi-hop retrieval. Figure 2 illustrates these limitations through concrete retrieval failure cases, providing clear motivation for the proposed approach. Although further analysis of their causes and frequency would strengthen the argument, the problem formulation is intuitive and well-grounded.

  • Simple and Intuitive Dual-path Propagation

NexusRAG introduces a corpus-level entity neighbor prior and incorporates it into semantic and structural propagation. Semantic propagation constrains entity activation using the neighbor prior alongside query–sentence similarity, while structural propagation explores neighboring entities without directly relying on query similarity. These complementary paths directly address the identified limitations. The neighbor prior is also reused for entity weighting during passage retrieval. The approach is intuitive and can be integrated into existing graph retrieval frameworks without additional relation extraction or complex learned models.

  • Consistent Improvements Across Datasets and Generation Backbones

The paper evaluates NexusRAG on three multi-hop QA benchmarks (HotpotQA, 2WikiMultiHopQA, and MuSiQue) and the Medical subset of GraphRAG-Bench, using three generation backbones: GPT-4o-mini, Qwen3.6-27B-FP8, and DeepSeek-V4-Flash. NexusRAG consistently outperforms LinearRAG across all evaluated settings. With GPT-4o-mini, LLM accuracy improves from 69.5% to 72.9% on HotpotQA, 65.0% to 68.7% on 2WikiMultiHopQA, 38.2% to 41.0% on MuSiQue, and 65.32% to 73.91% on Medical. These consistent gains across different generation backbones support the practical effectiveness of the proposed approach.

Critical weaknesses
  • Insufficient Analysis of the Proposed Failure Modes

The paper identifies query-gated cutoff and spurious activation as major retrieval failure modes but provides limited quantitative evidence of their prevalence and causes. Missing bridge evidence does not necessarily indicate query-gated cutoff, as failures may also originate from entity matching, activation pruning, passage initialization, or PPR ranking. Similarly, the actual contribution of spurious activation to retrieval failures remains unquantified.

  • Lack of Direct Validation of the Proposed Mechanisms

The authors claim that structural propagation mitigates query-gated cutoff and neighbor clamp suppresses spurious activation. However, the evaluation primarily focuses on downstream QA accuracy and retrieval recall. Although Appendices H and I provide supporting results and case studies, direct measurements are needed, such as the proportion of cutoff failures recovered by structural propagation and the reduction in irrelevant entity activations achieved by neighbor clamp.

  • Limited Retrieval Evaluation Despite Available Gold Evidence

Despite using three multi-hop QA benchmarks with gold supporting evidence, retrieval-level evaluation remains limited. In particular, All-Support Recall would help determine whether NexusRAG retrieves complete evidence chains rather than individual bridge or answer passages. Additionally, evaluating retrieval performance by hop count on MuSiQue would clarify its effectiveness for increasingly complex multi-hop reasoning.

  • Ablation Studies Do Not Fully Isolate Component Contributions

Although Table 4 demonstrates the importance of individual components, the ablations do not sufficiently disentangle their independent effects and interactions. For example, removing structural propagation still retains neighbor clamp and modified passage initialization, making performance attribution difficult. A factorial ablation study would provide stronger evidence of each component's contribution.

  • Insufficient Analysis of the Evidence Recall–Context Relevance Trade-off

NexusRAG improves Evidence Recall across all question types but consistently reduces Context Relevance. For Complex Reasoning, Evidence Recall increases from 87.03% to 94.62%, while Context Relevance drops from 81.58% to 68.52%. The authors attribute this to retrieving useful bridge evidence with low query similarity. However, the decline could also reflect additional irrelevant passages introduced through structural propagation. Further analysis is needed to distinguish useful intermediate evidence from retrieval noise, particularly given the paper's emphasis on suppressing spurious activation.

  • Unclear Relationship Between Evidence Coverage and QA Improvements

Retrieval improvements do not always align with downstream QA gains. For example, on 2WikiMultiHopQA, Answer-Recall@5 improves by only 0.4 percentage points, while Qwen3.6-27B-FP8 accuracy increases by 2.9 points. These gains may result from improved passage ranking or context composition rather than increased gold evidence coverage. Since NexusRAG modifies both entity propagation and passage initialization, the paper should disentangle these effects to explain the source of QA improvements.

  • Limited Generalizability Beyond LinearRAG

Although NexusRAG is compared against several GraphRAG baselines, its proposed mechanisms are evaluated only within LinearRAG. Their effectiveness in other relation-free GraphRAG frameworks remains unclear. Related methods such as EHRAG (Findings of ACL 2026), which combines entity co-occurrence and semantic information through structural–semantic hybrid diffusion, also warrant closer comparison. Furthermore, NexusRAG's mechanisms are closely tied to LinearRAG's entity–sentence–passage Tri-Graph. Without evaluating their transferability to other graph structures, the contribution is better characterized as a LinearRAG-specific enhancement rather than a broadly generalizable relation-free retrieval framework.

Questions for authors

See Weakness.

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

This paper proposes NexusRAG, an extension of LinearRAG, which is a relation-free graph RAG method that propagates entity activation over an entity–sentence–passage graph and ranks passages with Personalized PageRank. The authors identify two failure modes of LinearRAG's query-dependent propagation: useful bridge evidence with low query similarity can be cut off, and incidental entities in query-relevant sentences can be wrongly activated. To address them, NexusRAG builds a query-independent entity neighbor prior from co-occurrence and embedding similarity. It uses this prior to constrain the existing semantic propagation, to add a structural propagation along neighbor edges, and to initialize passage scores for PageRank. Experiments on three multi-hop QA benchmarks and a medical subset of GraphRAG-Bench show consistent gains in QA accuracy over LinearRAG and other graph-based RAG baselines.

Critical strengths
  1. The paper is well motivated. It starts from a concrete failure analysis of LinearRAG and identifies two specific failure modes, and the related work covers the main lines of graph-based RAG research.
  2. The method is simple and cheap. The neighbor prior is built without any LLM calls, the one-time precomputation is small, and the online retrieval time is essentially unchanged (Appendix J). This preserves the main efficiency advantage of relation-free graph RAG.
  3. The experiments are extensive. They cover four datasets, three generator backbones, re-judging with three evaluators, ablations of each use of the neighbor prior, parameter sensitivity, retrieval-level evaluation and case studies. NexusRAG's QA accuracy is higher than LinearRAG's on every dataset, with every backbone and under every evaluator.
Critical weaknesses
  1. The contribution is a targeted modification of LinearRAG: a query-independent neighbor prior that is used in its entity propagation and passage initialization. Propagating activation among entities that co-occur in sentences, and ranking passages with Personalized PageRank, are already part of LinearRAG, and the HippoRAG baselines also apply PageRank over entity graphs. The main addition is therefore explicit entity–entity neighbor edges built from co-occurrence and embedding similarity, and the paper would benefit from clarifying what is conceptually new beyond this addition.
  2. The evidence for the core mechanism could be more direct. The failure analysis in Section 3.1 is carried out only for LinearRAG, and for NexusRAG the two failure modes are only illustrated with two case studies. Apart from retrieval-level bridge recall (Appendix H), the paper does not directly measure whether NexusRAG reduces these failure modes during propagation, for example how often bridge entities are activated or incidental entities receive activation, compared with LinearRAG. Most of the evidence comes from final QA metrics.
  3. The evaluation relies mainly on two QA metrics (Contain-Acc and LLM-Acc), and only LLM-Acc is used for QA on the medical dataset; in the main table, the same model serves as both generator and judge, and the re-judging in Table 3 covers only LinearRAG and NexusRAG. Retrieval-level evidence is limited. On the multi-hop benchmarks, retrieval recall is reported only for HotpotQA and 2WikiMultiHopQA, and only against LinearRAG (Appendix H). On GraphRAG-Bench, the higher evidence recall comes with lower context relevance than LinearRAG in all four categories, which the abstract does not mention.
Questions for authors
  1. Could the authors provide a more direct comparison of the propagation itself with LinearRAG, for example by repeating the failure analysis of Section 3.1 for NexusRAG, or by measuring how often bridge entities and incidental entities are activated by each method?
  2. Could the authors report standard supporting-passage recall on all three multi-hop benchmarks, including MuSiQue, and explain why context relevance on GraphRAG-Bench is lower than LinearRAG's in all four categories?
  3. Were the baseline results re-run under the same settings as NexusRAG (for example, the same embedding model) or taken from prior work, and can the authors report significance tests for the main comparisons with LinearRAG?
Reviewer_BL6jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NexusRAG to address missing bridging evidence and incidental entity activation in graph retrieval-augmented generation without explicit relation extraction. Building on LinearRAG's entity-sentence-passage Tri-Graph, it introduces a query-independent neighbor prior that combines entity co-occurrence and semantic similarity. This prior constrains sentence-mediated semantic propagation and supports structural expansion between entities. The resulting entity evidence is also used to initialize passage scores, followed by Personalized PageRank for ranking. The authors evaluate the method on HotpotQA, 2WikiMultiHopQA, MuSiQue, and the Medical subset of GraphRAG-Bench, with component ablations and analyses of parameter sensitivity, retrieval recall, and efficiency. The results show fairly consistent improvements, although the overwrite update strategy and the specific design of the passage initialization formula require further justification.

Critical strengths

S1. The proposed approach shows consistent improvements across datasets and models, supported by substantial experimental evidence. Specifically, with GPT-4o-mini, NexusRAG improves Contain-Acc. over LinearRAG by 4.6, 2.7, and 3.2 percentage points on HotpotQA, 2Wiki, and MuSiQue, respectively. The improvements generally persist with additional generation models and across evaluators. The main experimental section and the appendices provide fairly comprehensive analyses of component ablations, parameter sensitivity, retrieval recall, and efficiency.

S2. The proposed approach achieves low additional computational cost and practical value. Specifically, the added neighbor structure does not rely on LLM-based relation extraction. Appendix J reports online retrieval times close to those of LinearRAG. On corpora of 5M and 10M tokens, neighbor precomputation accounts for approximately 2.6% and 2.8% of total indexing time, respectively. These results support the method's practical value as a lightweight improvement to an existing retrieval framework.

Critical weaknesses

W1. The overwrite update strategy in semantic propagation lacks sufficient justification. Section 3.2 states that when multiple admissible transitions reach the same entity, the latest assignment overwrites the previous score. Consequently, the propagation state in the next iteration may depend on the traversal order of entities or sentences. A weaker subsequent update may also overwrite a stronger one, potentially causing unintended effects. The authors do not sufficiently explain why overwriting is preferred to other aggregation strategies, such as summation or averaging, nor do they analyze the effect of traversal order on the results.

W2. The specific design of the passage initialization formula in Section 3.3 is insufficiently justified. Section 3.3 explains the intuitive role of each term in Eq. (4), but does not sufficiently justify the specific functional forms. First, why is an exponential transformation needed to preserve the dense retrieval ranking? The transformation changes the relative influence of the dense retrieval and entity evidence components; is this the motivation for using it? Second, the authors should explain how the scales of the two components are balanced when they are added directly, and whether either component may dominate. More targeted explanations are needed.

Questions for authors

Q1. What is the rationale for the overwrite update strategy? Does changing the traversal order of entities or sentences affect retrieval and QA results? Does this design offer consistent advantages over reasonable aggregation strategies?

Q2. What motivates the exponential transformation and the direct addition of the two components in Eq. (4)? Could the authors explain why these choices are needed and how the scales of the two components are balanced?

Minor clarifications and corrections:

M1. Appendix A.2 states that non-neighbor entities assigned a score of δ cannot participate in the next propagation step. However, there is a boundary case when σ_m = 1, which Appendix A.3 also acknowledges. Please ensure that the statements are consistent throughout the paper.

M2. The Related Work section describes Chain-of-Note as generating intermediate notes before retrieving supporting documents. The correct sequence is to retrieve documents first, then generate reading notes for the retrieved documents and use them to formulate an answer. Please correct this description.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.