acceptodds
Withdrawn from ICLR 2027

EVLINK: SOURCE-GROUNDED EVIDENCE LINKING FOR GRAPH RAG

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph-based Retrieval-Augmented Generation (GraphRAG) supports multi-hop reasoning by organizing corpora into structured graphs. However, graph reachabil- ity often captures semantic association rather than evidence support, so a reachable passage may still fail to justify a required cross-passage transition. We propose EvLink, an evidence-linking retriever that preserves passages as retrievable ev- idence units and builds evidence-supported transitions between them. EvLink constructs two types of reliable links: relation-grounded evidence links justified by explicit source relations, and endpoint-alignment links serving as source-bounded fallbacks. For retrieval, we introduce a two-stage retrieval strategy. First, bounded breadth-first search over source-grounded evidence links recovers bridge passages missed by similarity-based methods. Then, evidence-need mining with noisy-OR coverage refinement selects a compact, non-redundant evidence set satisfying distinct question facets. Experiments on three multi-hop and two simple QA bench- marks show EvLink consistently outperforms leading GraphRAG baselines with average gains of 2.4 R@5, 1.9 EM, and 2.4 F1.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Retrieval-Augmented Generation, GraphRAG, Multi-hop Question Answering, Evidence Retrieval, Source Grounding

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_yKZ5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes EvLink, a source-grounded evidence-linking framework for improving multi-hop retrieval in GraphRAG. The authors identify that existing GraphRAG methods often rely on semantic similarity or entity associations, which may introduce irrelevant passages and fail to recover complete evidence chains. To address this issue, EvLink constructs two types of cross-passage links: relation-grounded evidence links and endpoint-alignment links. It then employs bounded BFS search to retrieve candidate passages, followed by evidence-need mining and noisy-OR coverage optimization to select complementary and non-redundant evidence. Experiments on five question-answering benchmarks demonstrate that EvLink improves retrieval recall(R@5) , answer accuracy(EM) and F1 scores over existing RAG and GraphRAG baselines.

Critical strengths

1. Well-motivated and effective evidence-linking framework.

The paper addresses an important limitation of existing GraphRAG methods, where graph connectivity does not necessarily imply evidence support. By constructing source-grounded cross-passage links and combining them with evidence-need mining and coverage-based selection, EvLink provides a practical approach to improving multi-hop evidence retrieval.

2. Strong experimental results and comprehensive validation.

EvLink achieves competitive improvements over several recent GraphRAG and evidence-aware retrieval baselines across five QA benchmarks, particularly on 2WikiMultiHopQA. The ablation studies and controlled graph-structure experiments further demonstrate the effectiveness of the proposed components and show that the improvements cannot be explained solely by increased graph connectivity.

Critical weaknesses

1.Limited methodological novelty and insufficient comparisons with closely related work.

The proposed framework primarily combines existing techniques, including OpenIE-based relation extraction, graph traversal, question decomposition, and greedy coverage optimization. Although source-grounded evidence linking provides a meaningful distinction, the paper lacks direct comparisons with several recent approaches addressing similar challenges, such as LinearRAG, NeuroPath. This makes it difficult to establish the novelty and advantages of the proposed method over the latest GraphRAG techniques. Furthermore, a newly emerged method called SAG offers an approach that has also influenced the future sustainability of the research presented in this paper.

2.Insufficient validation of evidence-link reliability.

The paper claims that source-grounded links provide more reliable evidence connections, but the construction process mainly relies on extracted relations and entity matching, which do not necessarily guarantee valid evidential support. Although the ablation and structural control experiments demonstrate retrieval improvements, the paper lacks direct evaluations of link correctness and evidence-support quality. This weakens the justification for its central claim that the proposed links capture meaningful evidential relationships rather than merely improving graph connectivity.

Questions for authors

The authors propose a "Source-Grounded Evidence Links" method, the effectiveness of which is supported by extensive experiments and ablation studies. However, I believe the following aspects could be improved, or the authors should provide a rebuttal:

1. The contribution of the graph construction method is not sufficiently isolated

In the comparisons between EvLink and methods like HippoRAG 2, PropRAG, and NeocorRAG, not only do the graph construction approaches differ, but the subsequent retrieval, question decomposition, and re-ranking algorithms also vary. Consequently, even though EvLink achieves higher R@5, EM, and F1 scores, it remains unclear whether these improvements stem primarily from the "Source-Grounded Evidence Links" themselves or from the subsequent "Evidence-Need Mining" and "Noisy-OR" optimization. While Table 3 in the paper compares the method against Dense KNN, Same-seed KNN, and random graphs, these baselines do not constitute a direct comparison with state-of-the-art GraphRAG graph construction mechanisms. Therefore, while the current experiments demonstrate that the proposed graph structure outperforms certain simple graph structures, they do not sufficiently prove its superiority over advanced GraphRAG graph construction mechanisms.

2. There is a lack of direct evaluation regarding the correctness of the evidence links.

Although EvLink claims to construct "Source-Grounded Evidence Links," the actual mechanism relies primarily on: (OpenIE Fact+ Endpoint Matching)

For example: (Titanic, directed by, James Cameron)

Here, a cross-document link is established by matching the entity "James Cameron" to a target document. A potential issue arises here: the existence of a relational fact derived from a source does not guarantee that the resulting link actually supports the reasoning required for a given query. I suggest that the authors randomly sample a subset of the generated links and conduct human annotation to evaluate at least the following three metrics:

  • Relation Correctness: Whether the extracted relation is accurate.
  • Endpoint Alignment Accuracy: Whether the link correctly connects to the intended target document.
  • Evidence Support Validity: Whether the link genuinely provides meaningful cross-document evidence to support reasoning.

In particular, the correctness of "Relation-Grounded" links and "Endpoint-Alignment" links should be reported separately. This would provide direct evidence for the paper’s core claims, rather than merely demonstrating an improvement in final retrieval performance.

3.Experiments on complex multi-hop problems are not yet sufficient.

EvLink performs notably well on 2WikiMultiHopQA, but its improvement on MuSiQue is relatively limited. Furthermore, as shown in Appendix A.2 of the paper, the "evidence-need mining" component achieves an evidence recall rate of only 65% ​​for four-hop questions. It is recommended to provide a breakdown of results based on the number of reasoning hops—reporting metrics such as R@5, All@5, EM, and F1 separately for two-hop, three-hop, and four-hop questions. Additionally, the impact of the maximum number of hops in the BFS (breadth-first search) and the size of the candidate pool on the overall performance should be analyzed.

This would help determine whether EvLink is truly capable of handling complex multi-hop reasoning, rather than merely being effective at retrieving relatively simple chains of entity relations.

4.The research conclusions are stated too strongly

On page 6, the paper claims that EvLink outperforms all baseline methods across all benchmarks; however, Table 1 shows that on the MuSiQue dataset, the EM score for ETS-7B is 40.0%, while for EvLink it is 37.2%. Since ETS-7B and EvLink utilize different models for generating the final answer and the authors themselves have noted that a direct comparison is not appropriate.

If the author addresses these questions, I will consider raising the score.

Reviewer_7bB5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that edges in existing GraphRAG indices (HippoRAG/HippoRAG 2, PropRAG, HGRAG) capture topical association rather than evidential support, and proposes EvLink, a passage-level graph whose edges are licensed by OpenIE facts extracted from the source passage whose endpoint string-matches another passage's title or alias. Edges are split into "relation-grounded" links (the fact has a non-empty relation phrase) and "endpoint-alignment" fallback links. At query time, dense top-20 seeds plus title/anchor matches start a 2-hop bounded BFS (cap 200 passages), and a second stage uses an LLM to decompose the question into 1-4 evidence requirements B(q), runs an LLM-based "dependency binding" step that reads retrieved passages to resolve bridge entities, and then picks the top-5 passages by greedily maximizing a noisy-OR coverage objective (with the standard (1-1/e) guarantee in Appendix A.4). On the HippoRAG 2 1,000-query splits of HotpotQA, 2Wiki, MuSiQue, NQ and PopQA with NV-Embed-v2, a Qwen3-32B extractor and a GPT-4o-mini reader, EvLink is reported best on R@5/EM/F1 against reproduced baselines (Table 1), with the largest gain on 2WikiMultiHopQA (+5.3 R@5 over PropRAG). Ablations (Table 2), graph-substitution controls (Table 3), a new FCRG diagnostic (Table 4), token cost (Table 5) and extractor-swap experiments (Table 6, Table 15) are also provided.

Critical strengths
  1. The baselines are mostly run under one controlled protocol: the same Qwen3-32B for every LLM-assisted indexing/question step, the same NV-Embed-v2 encoder, the same GPT-4o-mini reader with an identical prompt, the same top-5 context and the same answer normalization (Appendix A.1, A.3). The authors are also upfront that ETS and Relink are copied from the original papers with different readers (Appendix A.3). I checked the arithmetic in the text against Table 1 (e.g., the 2Wiki gaps over NV-Embed-v2, RAPTOR, PropRAG, HippoRAG 2 and HGRAG, and the NeocorRAG multi-hop averages 85.7/53.0/63.1 vs. 89.1/55.3/66.5), and it all matches.

  2. The paper tries to go beyond a plain component ablation. Table 3 swaps the evidence-link graph for dense-kNN, edge-count-matched, same-seed and degree-matched shuffled graphs, Table 10 separates relation-grounded from endpoint-alignment edges, and Appendix A.2 includes a manual audit of B(q) quality with exact CIs and a per-hop breakdown for MuSiQue (94% / 83% / 65% answer-sufficient on 2/3/4 hops). That breakdown is honest and useful, since it shows exactly where decomposition breaks down.

  3. The cost reporting is more complete than is typical for GraphRAG papers: Table 13 has exact offline OpenIE token counts per dataset, Table 14 breaks down online per-query auxiliary tokens, Table 12 reports end-to-end retrieval latency, and the extractor-swap experiments (Qwen3-8B in Table 6, Llama 3.3 70B in Table 15) show that the method's ranking holds when the upstream LLM changes. The idea of keeping passages as the retrieval unit and attaching the provenance span to each edge is simple and practical.

Critical weaknesses
  1. Most of the gain seems to come from query-time LLM decomposition and binding rather than from the evidence links. On 2Wiki, "w/o source-grounded evidence links" still reaches 92.9 R@5 / 81.0 All@5 (Table 2), up from 49.3 dense All@5 (Table 9), so the links add only about +8.5. Both cases in Table 11 are also rescued by binding, so the link-centric framing should be revised, and binding needs to move out of the appendix.

  2. Decomposition and IRCoT-style baselines at a matched LLM-call budget are missing. The reproduced baselines also look weaker than their published numbers: Table 15's Llama rerun lifts HippoRAG 2 to 95.3/89.8/73.0 R@5. Outside 2Wiki, the margins on HotpotQA and MuSiQue are 0.1-1.0 EM/F1 with no CIs or significance tests, and the abstract's 2.4/1.9/2.4 averages cover only the three multi-hop datasets.

  3. Every control graph in Table 3 scores below having no graph at all (2Wiki All@5 is 69.7 for kNN and 51.9 for shuffled, vs. 81.0 without links in Table 2). That suggests the controls degrade the candidate pool rather than showing that edge semantics matter. FCRG (Table 4) also looks circular, since its audit set appears to be defined by EvLink's own OpenIE transitions.

Questions for authors
  1. Can you report a strong decomposition baseline at a matched budget: NV-Embed-v2 + your B(q) + dependency binding + coverage selection over a dense-only Cq (no graph, no anchors), plus IRCoT on top of HippoRAG 2/PropRAG? How does this relate to the "w/o source-grounded evidence links" row? Does that row keep question-anchor seeding D_A(q) and dense fill to 100?

  2. Why does every Table 3 control graph (kNN, edge-count, shuffled) do worse than removing the graph entirely (Table 2)? Were Cq and L constructed identically across these variants, and what fraction of Cq is dense-filled vs. graph-admitted in each?

Reviewer_n7SnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduce EvLink, which is an approach to GraphRAG, where evidence linkages are generated based on source-based evidences for establishing relationships between passages. In the offline stage, the OpenIE triples in the passage are aligned against the title of other passages, creating directed relation-based or endpoint alignment links. Then, bridge passages are generated using BFS on these links and selection of evidence is done through question decomposition/noisy-OR coverage-based selection. The experiments have been performed on three multi-hop and two simple QA datasets achieving an improvement of 2.4 R@5, 1.9 EM, and 2.4 F1 over baselines.

Critical strengths
  1. This paper tackles a real-world problem. While graph-connected passages may not contain the information necessary to answer the question, keeping source passages and text spans linked to links preserves these connections for examination.

  2. This paper evaluates models for both multi-hop and single-hop QA problems. Under the common framework setting, EvLink exhibits good performance, especially in 2WikiMultiHopQA.

  3. The analysis extends the findings presented. Graph controls from Table 3, human inspection of question decomposition, and experiments with different upstream LLMs all contribute to a better understanding of the system.

Critical weaknesses
  1. Linking requires that the ending of a fact in the source passage matches the title (or title variations) of the target passage. This presupposes a corpus where each passage is an entity page having a unique and identifiable title. In those corpora which have no titles, numerous passages which share the same title, or title-like passages (such as enterprise documents, technical manuals, news, scientific papers, etc.), very few links would be constructed and EvLink degenerates into dense retrieval and re-ranking. All five corpora are Wikipedia-based having gold titles.

  2. The improvements are mostly seen in 2Wiki, without any statistical significance. The improvements compared to the best baseline model are +0.5 in F1 for HotpotQA, +0.1 EM / +1.0 F1 for MuSiQue, and +0.3 F1 for NQ. Because the questions from 2Wiki are generated using the hops from entity-to-page, these questions favor the construction of title links, and thus the improvement is mostly because of the structure of the dataset.

  3. Several of the approaches referenced in this paper have not been compared to the proposed approaches, such as LineageRAG [1], PathRAG [2], SentGraph [3], and GraphRAG [4]. Also missing from citation and comparison are the following highly relevant papers: HopRAG [5], KG²RAG [6], SiReRAG [7], GFM-RAG [8], LinearRAG [9], LightRAG [10], graph-based recurrent retrieval on Wikipedia [11], and IRCoT [12].

[1] LineageRAG: Harnessing GraphRAG by constructing evidence lineages with source grounding.

[2] PathRAG: Pruning graph-based retrieval augmented generation with relational paths.

[3] SentGraph: Hierarchical sentence graph for multi-hop retrieval-augmented question answering.

[4] From local to global: A graph RAG approach to query-focused summarization.

[5] HopRAG: Multi-hop reasoning for logic-aware retrieval-augmented generation.

[6] Knowledge graph-guided retrieval augmented generation.

[7] SiReRAG: Indexing similar and related information for multihop reasoning.

[8] GFM-RAG: Graph foundation model for retrieval augmented generation.

[9] LinearRAG: Linear graph retrieval augmented generation on large-scale corpora.

[10] LightRAG: Simple and fast retrieval-augmented generation.

[11] Learning to retrieve reasoning paths over Wikipedia graph for question answering.

[12]Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.

  1. Unfair baseline comparison that impacts the claims made. The numbers provided by ETS and Relink are taken from their own papers, where they use different models (Qwen2.5-72B-Instruct and DeepSeek-V3-0324) and likely different data splits and evaluation protocols, but presented in the same main table as directly comparable results. Such information is only provided in the Appendix A.3, while the Table 1 footnote states only that R@5 is not reported. Moreover, ETS reaches 40.0 EM on MuSiQue compared to 37.2 EM of EvLink, refuting the claim that “EvLink consistently outperforms all baselines across every benchmark.”

  2. Some important experiments are missing. (1) No hyperparameter sensitivity analysis. The Table 8 considers just one set of parameters for all benchmarks, and no parameter that directly influences the linking and recall processes is considered. (2) Just one reader (GPT-4o-mini) and one embedder (NV-Embed-v2) are used.

  3. The incomplete method description provided in the main text refers to "surface mentions," but the description in Appendix A.3 refers exclusively to title variations. Table 8 features a "synonymy similarity threshold" set at 0.85, with no explanation of its purpose.

Questions for authors

1.Could you evaluate EvLink on corpora not from Wikipedia or with titles stripped off passages? How does the number of links generated and performance change?

  1. Could you provide graph statistics such as average out-degree, proportion of isolated passages, |E_rel| and |E_end|, and their mean/std significance testing for multiple runs?

  2. Could you compare EvLink to related approaches like LineageRAG, HopRAG, and KG²RAG, etc.?

  3. Could you provide the sensitivity analysis for parameters h, L, number of seeds, hub-degree limit, and K, as well as the MuSiQue statistics for different hops (2/3/4)?

  4. What is the meaning of the synonymy threshold (0.85) in Table 8 and how do you get "surface mentions"?

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

Related papers

Discussion (0)

Sign in to comment.