EvLink: Source-Grounded Evidence Linking for Graph RAG

arXiv:2609.29695 · cs.IR · Submitted 2026-08-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "EvLink: Source-Grounded Evidence Linking for Graph RAG".

Jane: Here's the gist of the paper. Graph-based Retrieval-Augmented Generation (GraphRAG) supports multi-hop reasoning by organizing corpora into structured graphs However, graph reachability often captures semantic association rather than evidence support, so a reachable passage may still fail to justify a required cross-passage transition The paper proposes EvLink, an evidence-linking retriever that preserves passages as retrievable evidence 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 sourcebounded fallbacks For retrieval, a two-stage strategy is introduced: first, bounded breadth-first search over source-grounded evidence links recovers bridge passages missed by similarity-based methods Then, evidenceneed 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 benchmarks show EvLink consistently outperforms leading GraphRAG baselines with average gains of 2.4 R@5, 1.9 EM, and 2.4 F1 Evidence Link Construction EvLink constructs a substrate over source passages while preserving passages as the fundamental retrievable units It constructs two types of links: relation-grounded evidence links connect a source passage to a target passage via an explicit source relation and an aligned endpoint An endpoint-alignment link is proposed when an endpoint of a source-grounded fact in di aligns to dj but no relation-grounded transition is available The final edge set is defined as E = Erel ∪ Eend, where Erel are relation-grounded links and Eend are endpoint-alignment links Two-Stage Retrieval Strategy For query-conditioned retrieval, EvLink employs a two-stage strategy First stage involves dense entries and question anchors to localize entry passages, followed by bounded breadth-first search (BFS) to induce a query-local candidate pool that recovers bridge passages missed by similarity-based methods The second stage uses evidence-need mining and noisy-OR coverage refinement to select a compact evidence set satisfying distinct answer facets while eliminating traversal-order redundancy Performance Gains and Mechanism Analysis EvLink consistently outperforms leading baselines across every benchmark, demonstrating its superior effectiveness in QA tasks The most pronounced performance gap emerges on 2WikiMultiHopQA, where EvLink surpasses the NV-Embed-v2 by 20.3 R@5, 10.9 EM and 14.2 F1 points The improvement arises because EvLink ranks passages through source-grounded evidence links instead of relying on graph reachability alone Ablation studies show that stripping away relation-grounded links causes a noticeable decline in performance, particularly on 2WikiMultiHopQA, dropping by 3.3 points in R@5 and 8.5 points in All@5 Efficiency and Adaptability EvLink uses an average of 13.36M auxiliary tokens per 1,000-query multi-hop evaluation This total is lower than HippoRAG 2 (14.12M), PropRAG (18.84M), and NeocorRAG (19.78M) among leading graph-based RAG baselines EvLink ranks first in R@5, EM, and F1 with both Qwen3-8B and Qwen3-32B The advantage persists even with the smaller upstream model Retrieval Quality Criteria Instantiation The evidence-need mining step defines the query-side evidence needs used to select from Cq The noisy-OR coverage objective maximizes coverage of the evidence requirements B(q) by selecting passages that satisfy distinct answer facets while eliminating traversal-order redundancy This process ensures bridge and tail passages are retained only if they provide non-redundant coverage of unmet evidence requirements The mechanism demonstrates that source-grounded document transitions are decisive for recovering this evidence class EvLink presents a novel GraphRAG framework that builds evidence links between source passages By keeping passages as the basic retrieval unit and building relation-grounded links only when supported by explicit textual justification, our method avoids the high cost of finegrained semantic decomposition while enabling source-grounded cross-passage transitions The two-stage evidence-conditioned retrieval pipeline further aligns graph structure with question needs through bounded BFS for bridge recovery and noisy-OR coverage refinement to remove redundancy The superior results of EvLink across five QA benchmarks validate that source-grounded evidence links can improve GraphRAG retrieval by supporting cross-passage evidence recovery Limitations The present study focuses on established multi hop QA benchmarks, which provide a controlled setting for evaluating evidence retrieval over passage collections Extending EvLink to this setting requires further study A promising direction is iterative retrieval in which accumulated evidence guides subsequent evidence link traversal through the document structure Ethical Considerations The main ethical risk comes from unreliable information in the underlying corpora, which may lead the reader to produce unsupported answers Source grounded transitions improve auditability without guaranteeing factual correctness Sensitive applications therefore require expert review of the retrieved evidence References Gabor Angeli, Melvin Jose Johnson Premkumar, and Christopher D.

Tom: Manning.

Jane: 2015 Boyu Chen, Zirui Guo, Zidan Yang, Yuluo Chen, Junze Chen, Zhenghao Liu, Chuan Shi, and Cheng Yang.

Tom: 2026 Darren Edge et al.

Paper summary: Tom: We've gone through "EvLink: Source-Grounded Evidence Linking for Graph RAG" and seen how they build those links based on explicit source relations versus just general graph reachability. The authors are very clear that this evidence linking approach provides superior performance across the benchmarks they tested, especially where simple reachability fails.

Jane: It seems the main contribution of this paper is shifting the focus from just organizing data into a graph to building an evidence-linked substrate over those passages, ensuring that any connection made between them is supported by actual textual justification.

Lu: The implication for future work they suggest is iterative retrieval, where accumulated evidence actually guides subsequent link traversal through the document structure in a more guided way. That suggests we can build systems that learn how to explore based on what they've already found, not just static graph structures.

Meng: I worry about extending this to real traffic. If the cost per query is manageable and it keeps recovering bridge passages effectively, then maybe this is something worth exploring for our next generation of reasoning models.

Tom: It sounds like the core message of EvLink is that you need source grounded evidence links to improve Graph RAG retrieval by supporting cross passage evidence recovery, which they validated with results on five different QA benchmarks.

Conclusion: Tom: So we're wrapping up on EvLink, this paper by Linyao Zheng and Xuhang Shi, and what they've done with Graph RAG. The big idea here is building evidence links between passages rather than just connecting them randomly in a graph structure.

Jane: Right, so instead of just saying two documents are vaguely related because they share some topic, EvLink builds these links only when there’s actual text in the source material that justifies the connection. It's about making sure those connections have proof attached to them.

Lu: Exactly. They introduce these two types of links—relation grounded evidence links and endpoint alignment links—to make sure the system doesn't just follow any path, but only paths backed by explicit facts in the text.

Meng: So, if we think about what this means for deploying stuff at scale, it’s that we’re moving away from systems that rely on fuzzy graph reachability to something much more traceable. It means you can actually audit *why* the AI picked a certain answer based on the evidence chain.

Tom: And the results show this matters. They beat leading Graph RAG baselines on several benchmarks, with gains like two point four R at five and two point four F one points across different tests. That’s concrete performance improvement we can talk about.

Jane: But what does that actually mean for someone just listening? It means that when you ask a complex question, the system doesn't just guess based on a broad neighborhood; it follows the specific textual evidence trail to find the right information.

Tom: It changes how we think about retrieval in these massive knowledge graphs. Instead of hoping a path exists, we’re building a structure where paths only exist if they are supported by documented relations between pieces of text.

Lu: And that two-stage retrieval strategy—using bounded breadth first search to find those missing bridges and then using evidence need mining to pick the best set of evidence—that’s the technical trick they used to make it work on those hard multi-hop questions.

Meng: It's still a bit heavy on tokens, though. They use about thirteen point three million auxiliary tokens per thousand queries, which is less than some other big graph RAG systems, but that’s a cost we have to watch when we try to ship this into production environments.

Jane: So while it’s much better at finding the right evidence for hard questions, they also flag that they are still focusing on controlled benchmarks and suggest more work is needed for real-world, iterative retrieval in documents.

Tom: Right. So EvLink shows that making connections evidence-based, not just graph-based, gives you a tangible lift in accuracy across the board when dealing with multi hop questions.

Jane: And that leads us to thinking about what comes next—how we can take this idea and apply it to documents where the structure isn't as neat as these tests suggest.

Linyao Zheng, Xuhang Shi, Zhifang Mao, Sai Zhou, Shuaixian An, Xiuquan Hou

China Mobile Communications Group Shaanxi Co., Ltd. · Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University

cs.IR

Submitted: 2026-08-31

Updated: 2026-11-08

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: Graph-based Retrieval-Augmented Generation (GraphRAG) supports multi-hop reasoning by organizing corpora into structured graphs However…

Key concepts

Relation-Grounded Evidence Links
These are specific connections made between passages only when there is an explicit source relation linking them. This means the link isn't just based on general similarity; it must be justified by a defined relationship within the source documents, making the evidence more reliable for multi-hop reasoning.
Endpoint-Alignment Links
These links are used as a fallback when a direct relation-grounded transition is missing. They connect passages if an endpoint of one fact aligns with an endpoint of another fact, serving as a necessary connection to complete the evidence chain in complex queries.
Noisy-OR Coverage Refinement
This is a technique used during the final retrieval stage to select the best set of evidence. It maximizes coverage of all required answer facets while actively removing redundant passages that provide overlapping information, ensuring a compact and unique set of supporting evidence is chosen.

Terminology

Summary

Graph-based Retrieval-Augmented Generation (GraphRAG) supports multi-hop reasoning by organizing corpora into structured graphs However, graph reachability often captures semantic association rather than evidence support, so a reachable passage may still fail to justify a required cross-passage transition The paper proposes EvLink, an evidence-linking retriever that preserves passages as retrievable evidence 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 sourcebounded fallbacks For retrieval, a two-stage strategy is introduced: first, bounded breadth-first search over source-grounded evidence links recovers bridge passages missed by similarity-based methods Then, evidenceneed 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 benchmarks show EvLink consistently outperforms leading GraphRAG baselines with average gains of 2. p. 04 R@5, 1.9 EM, and 2.4 F1

Evidence Link Construction

EvLink constructs a substrate over source passages while preserving passages as the fundamental retrievable units It constructs two types of links: relation-grounded evidence links connect a source passage to a target passage via an explicit source relation and an aligned endpoint An endpoint-alignment link is proposed when an endpoint of a source-grounded fact in di aligns to dj but no relation-grounded transition is available The final edge set is defined as E = Erel ∪ Eend, where Erel are relation-grounded links and Eend are endpoint-alignment links

Two-Stage Retrieval Strategy

For query-conditioned retrieval, EvLink employs a two-stage strategy First stage involves dense entries and question anchors to localize entry passages, followed by bounded breadth-first search (BFS) to induce a query-local candidate pool that recovers bridge passages missed by similarity-based methods The second stage uses evidence-need mining and noisy-OR coverage refinement to select a compact evidence set satisfying distinct answer facets while eliminating traversal-order redundancy

Performance Gains and Mechanism Analysis

EvLink consistently outperforms leading baselines across every benchmark, demonstrating its superior effectiveness in QA tasks The most pronounced performance gap emerges on 2WikiMultiHopQA, where EvLink surpasses the NV-Embed-v2 by 20.3 R@5, 10.9 EM and 14.2 F1 points The improvement arises because EvLink ranks passages through source-grounded evidence links instead of relying on graph reachability alone Ablation studies show that stripping away relation-grounded links causes a noticeable decline in performance, particularly on 2WikiMultiHopQA, dropping by 3. p. 53 points in R@5 and 8. p. 55 points in All@5

Efficiency and Adaptability

EvLink uses an average of 13.36M auxiliary tokens per 1,000-query multi-hop evaluation This total is lower than HippoRAG 2 (14. p. 712M), PropRAG (18. p. 784M), and NeocorRAG (19. p. 778M) among leading graph-based RAG baselines EvLink ranks first in R@5, EM, and F1 with both Qwen3-8B and Qwen3-32B The advantage persists even with the smaller upstream model

Retrieval Quality Criteria Instantiation

The evidence-need mining step defines the query-side evidence needs used to select from Cq The noisy-OR coverage objective maximizes coverage of the evidence requirements B(q) by selecting passages that satisfy distinct answer facets while eliminating traversal-order redundancy This process ensures bridge and tail passages are retained only if they provide non-redundant coverage of unmet evidence requirements The mechanism demonstrates that source-grounded document transitions are decisive for recovering this evidence class

Conclusion

EvLink presents a novel GraphRAG framework that builds evidence links between source passages By keeping passages as the basic retrieval unit and building relation-grounded links only when supported by explicit textual justification, our method avoids the high cost of finegrained semantic decomposition while enabling source-grounded cross-passage transitions The two-stage evidence-conditioned retrieval pipeline further aligns graph structure with question needs through bounded BFS for bridge recovery and noisy-OR coverage refinement to remove redundancy The superior results of EvLink across five QA benchmarks validate that source-grounded evidence links can improve GraphRAG retrieval by supporting cross-passage evidence recovery

Limitations

The present study focuses on established multi hop QA benchmarks, which provide a controlled setting for evaluating evidence retrieval over passage collections Extending EvLink to this setting requires further study A promising direction is iterative retrieval in which accumulated evidence guides subsequent evidence link traversal through the document structure

Ethical Considerations

The main ethical risk comes from unreliable information in the underlying corpora, which may lead the reader to produce unsupported answers Source grounded transitions improve auditability without guaranteeing factual correctness Sensitive applications therefore require expert review of the retrieved evidence

References

Gabor Angeli, Melvin Jose Johnson Premkumar, and Christopher D. Manning. 2015 Boyu Chen, Zirui Guo, Zidan Yang, Yuluo Chen, Junze Chen, Zhenghao Liu, Chuan Shi, and Cheng Yang. p. 8 2026 Darren Edge et al. p. 8 p. 8 2024 Wenqi Fan et al. p. 8 p. 8 2024 Bernal Jiménez Gutiérrez et al. p. 8 p. 8 2024 Curran Associates, Inc. p. 8 p. 8

--- Page 1 ---

EvLink: Source-Grounded Evidence Linking for Graph RAG

Linyao Zheng, Xuhang Shi, Zhifang Mao, Sai Zhou, Shuaixian An, Xiuquan Hou Correspondence: Linyao Zheng (zhenglinyao@sn.chinamobile. p. 0com), Xuhang Shi (xuhang. p. 0shi@outlook. p. 0com)

--- Page 2 ---

Figure 1: GraphRAG failures vs. EvLink Topical reachability objectives cause failure in GraphRAG: retrieval remains in the local movie neighborhood, showing that GraphRAG captures neighborhood relevance In contrast, EvLink uses source-grounded evidence links to admit required passage transitions and close the evidence chain

--- Page 3 ---

Method: EvLink

EvLink has three stages: Evidence Link Construction builds source-grounded inter-passage links while preserving passages as retrievable units Coarse-Grained Evidence Induction seeds dense entries and question anchors, then runs bounded BFS over evidence links Fine-Grained Evidence Retrieval mines evidence needs and selects compact support via noisy-OR coverage

--- Page 4 ---

Figure 2: Overview of EvLink The query-local subgraph keeps the seed documents and all evidence links incident to the bounded-hop neighborhood

--- Page 5 ---

Table 1: Main results EvLink achieves superior performance over baselines Compared with LLM-based retrieval and conventional RAG, the most pronounced performance gap emerges on 2WikiMultiHopQA

--- Page 6 ---

Table 1: Main results Compared with Graph RAG baselines, PropRAG is the strongest competitor on 2WikiMultiHopQA, but EvLink still improves over it by 5.3 R@5 points, 4. p. 58 EM points, and 5. p. 56 F1 points

--- Page 7 ---

Table 1: Main results Compared with evidence-aware RAG baselines, NeocorRAG is the strongest directly comparable system

--- Page 8 ---

Table 3: Mechanism Analysis Dense similarity alone is not enough EvLink reaches 0.300 FCRG and recovers 67. p. 66% of audited documents

--- Page 11 ---

Table 11: Case studies on 2WikiMultiHopQA showing how evidence-need mining and coverage refinement recover bridge passages missed by dense and traversalonly retrieval

--- Page 13 ---

Table 8: Hyperparameters used in all EvLink experiments, including the main table and ablations Table 8 lists all hyperparameters used in the main experiments, including the main table and ablations

--- Page 14 ---

Table 7: Correct cases in the manual audit of generated evidence requirements Table 7 shows the corresponding system effect

--- Page 15 ---

Table 13: Offline generative-token cost for corpus structuring Table 13 breaks down generative-token consumption

--- Page 16 ---

Table 14: Online auxiliary generative-token cost per query, excluding the shared GPT-4o-mini reader Table 14 breaks down generative-token consumption

--- Page 17 ---

Table 8: Hyperparameters used in all EvLink experiments, including the main table and ablations Table 8 lists all hyperparameters used in the main experiments, including the main table and ablations

--- Page 18 ---

Table 15: Recall@5 under two upstream language models using a shared evaluation protocol Table 15 shows average Recall@5 gains of 3.91 points for HippoRAG 2, 1. p. 517 points for PropRAG, and 1. p. 563 points for EvLink with Llama 3. p. 53 70B Instruct

--- Page 8 ---

Table 4: FCRG and All@5 recovery Table 4 gives the same conclusion on dense-missed fact-conditioned evidence The gap shows that recovering this evidence class depends on source-grounded document transitions, not matched connectivity statistics

--- Page 13 ---

Table 14: Online auxiliary generative-token cost per query, excluding the shared GPT-4o-mini reader Table 14 breaks down generative-token consumption The requirement-decomposition tokens are estimated with the cl100k base tokenizer from the stored prompts and saved requirement outputs The binding tokens are exact usage records from the binding calls

--- Page 14 ---

The candidate with the highest marginal gain is selected: dt = arg max d∈Cq∖Rt-1 ∆q(d Rt-1) The process terminates when Rt = K or Cq is exhausted, yielding Rq = Rt

--- Page 15 ---

The process terminates when Rt = K or Cq is exhausted, yielding Rq = Rt The final stage targets the constrained objective R⋆ q = arg max R⊆Cq R≤K Covq(R)

Gabor Angeli, Melvin Jose Johnson Premkumar, and Christopher D. Manning. 2015 Boyu Chen et al. p. 8 2026 Darren Edge et al. 2024 Wenqi Fan et al. 2024 Bernal Jiménez Gutiérrez et al. 2024 Curran Associates, Inc.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Relation-grounded evidence links construct transitions based on explicit source relations and endpoint-alignment links serving as a fallback through source-bounded endpoint alignment, ensuring that graph connectivity is strictly contingent on textual evidence. This shifts graph connectivity from topical reachability to evidential necessity.

  2. The two-stage retrieval strategy allows for localization followed by recovery of bridge passages: the first stage uses bounded breadth-first search over source-grounded evidence links recovers bridge passages missed by similarity-based methods, and the second stage uses evidence-need mining with noisy-OR coverage refinement selects a compact, non-redundant evidence set.

  3. The system can recover missing bridge passages through Fact Conditioned Rank Gain (FCRG), which measures the gain from gold evidence that is reachable from another gold support passage through a fixed source-grounded fact transition, demonstrating that source-grounded document transitions are necessary for recovering evidence class.

  4. The system can select non-redundant, question-side requirements by using the noisy OR coverage objective to maximize total coverage utility, ensuring that bridge and tail passages are retained only if they provide non-redundant coverage of unmet evidence requirements.

  5. The system exhibits high adaptability across different base LLMs; for instance, EvLink ranks first in R@5, EM, and F1 with both Qwen3-8B and Qwen3-32B, showing that the advantage persists even with the smaller upstream model.

Sources

Related papers