acceptodds
Withdrawn from ICLR 2027

DAG-RAG: Multi-Path Hierarchical Entity Grounding for GraphRAG

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Retrieval-augmented generation (RAG) enhances large language models by retrieving external knowledge, and GraphRAG further enhances this paradigm by organizing knowledge as graphs for structured reasoning. However, existing GraphRAG methods still struggle with grounding complex queries to relevant entities or subgraphs within a large graph space, leading to inefficient retrieval and unstable downstream generation. To address these issues, we propose DAG-RAG, a hierarchical GraphRAG framework that constructs a multi-granularity directed acyclic graph from the original knowledge graph and performs coarse-to-fine grounding with a path-conditioned router. The structured design also enables incremental maintenance by updating only affected clusters without full graph reconstruction or model retraining Experiments show that DAG-RAG improves retrieval accuracy and downstream answer quality while reducing unnecessary graph traversal, demonstrating the effectiveness of hierarchical DAG grounding for scalable and adaptive GraphRAG.

Primary area:
foundation or frontier models, including LLMs
Keywords:
retrieval-augmented generation, graph retrieval-augmented generation, entity grounding, hierarchical retrieval, directed acyclic graph index, learned router, multi-hop question answering

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_v95fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DAG-RAG, a hierarchical GraphRAG framework designed to improve entity grounding for complex queries. It constructs a multi-parent DAG through hierarchical clustering and introduces a path-conditioned router that progressively selects relevant entities from coarse to fine. By aggregating probabilities across multiple paths, DAG-RAG aims to reduce early-stage grounding errors while controlling retrieval costs. Experiments on HotpotQA, MuSiQue, and PopQA demonstrate competitive answer quality and retrieval efficiency.

Critical strengths
  1. The paper identifies entity grounding as a critical bottleneck in GraphRAG, where incorrect seed entities can undermine subsequent retrieval and reasoning.

  2. The multi-parent DAG, path-conditioned routing, and probability aggregation form a well-structured framework for coarse-to-fine entity grounding.

  3. DAG-RAG achieves strong results across three QA benchmarks while requiring fewer LLM tokens than most GraphRAG baselines.

Critical weaknesses
  1. This framework primarily integrates existing technologies, including hierarchical clustering, multi-parent indexing, GRU-based path modeling, and probabilistic aggregation. However, what exactly are its advantages over existing hierarchical retrieval methods? Is path modeling the key factor?

  2. DAG-RAG fine-tunes its entity encoder and trains the router using oracle entity labels, while the main baselines do not receive equivalent supervision. This makes it difficult to attribute the improvements to the proposed architecture.

  3. The paper does not adequately demonstrate why recurrent path modeling is necessary compared with simpler parent-child scoring or learned hierarchical routing. Existing ablations do not fully isolate this contribution.

  4. Although incremental updates are highlighted as an advantage, the paper only reports a single update-time measurement (21.8 seconds), excluding LLM-based entity and relation extraction. It lacks comparisons with full reconstruction, evaluations of retrieval accuracy after updates, and experiments on repeated updates. Therefore, the claimed efficiency and reliability of incremental maintenance remain insufficiently supported.

  5. This router relies on oracle entity labels automatically constructed from annotated supporting evidence; however, how reliable are these labels? Their quality should directly determine the router’s effectiveness. Furthermore, zero-shot transfer has only been tested on PopQA, a dataset that shares similar domain characteristics with the training benchmark. This raises concerns about whether the model can generalize to unseen domains in the absence of real entity supervision.

Questions for authors

See weaknesses above.

Reviewer_9Pu5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DAG-RAG, a hierarchical GraphRAG framework that addresses the challenge of reliably matching complex queries to relevant knowledge graph entities while improving graph traversal efficiency. DAG-RAG builds a multi-parent hierarchical structure through semantic clustering and introduces a path-conditioned router for coarse-to-fine entity retrieval.

Critical strengths
  1. The paper is well motivated, and tackles an important problem in GraphRAG. Existing methods relying on embedding similarity or exact matching often struggle to identify relevant entities when queries are complex or expressed differently from the entities in the graph.
  2. The idea of combining hierarchical retrieval with path-conditioned routing is good. Allowing multiple paths to reach the same entity seems like a reasonable way to reduce the impact of incorrect early routing decisions.
Critical weaknesses
  1. Baselines are outdated, mainly from 2024 to 2025. This is an ICLR 2027, I would expect comparison with more recent methods such as G-Reasoner, Youtu-GraphRAG [2] and Graph-r1 [3].
  2. The proposed hierarchy is built purely on semantic similarity without considering the actual relations in the knowledge graph. This introduces an additional routing stage before expanding the original knowledge graph, making the overall retrieval process more complex without fully leveraging the graph structure
  3. The paper lacks comparisons with query expansion/rewriting methods, which are commonly used to address ambiguous or complex queries by leveraging LLMs to generate more informative queries or entities.

[1] Luo, L., Zhao, Z., Liu, J., Qiu, Z., Dong, J., Panev, S., ... & Pan, S. (2026). G-reasoner: Foundation models for unified reasoning over graph-structured knowledge. ICLR [2] Dong, J., An, S., Yu, Y., ZHANG, Q., Luo, L., Huang, X., ... & Sun, X. (2026). Youtu-graphrag: Vertically unified agents for graph retrieval-augmented complex reasoning. ICLR [3] Luo, H., Chen, G., Lin, Q., Guo, Y., Xu, F., Kuang, Z., ... & Tuan, L. A. (2026). Graph-r1: Towards agentic graphrag framework via end-to-end reinforcement learning. ICML

Questions for authors
  1. Since the proposed hierarchy is built purely on semantic similarity, why not incorporate the actual knowledge graph relations into the hierarchical construction? Would this help better capture relational information while avoiding the additional traversal complexity?

  2. Could the authors include comparisons with query rewriting and expansion techniques, as well as other hierarchical or schema-based GraphRAG methods? This would help demonstrate whether the proposed approach offers advantages over alternative solutions to the same entity-grounding problem.

Reviewer_K4boOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses entity grounding in GraphRAG, arguing that retrieval errors often originate before subgraph expansion, when a query is mapped to the wrong entities. DAG-RAG builds a multi-parent hierarchical index over entity embeddings and traverses it through a learned, path-conditioned router. The router combines local relevance scores with information accumulated along different paths, using a candidate budget to limit the search. The resulting entities are expanded and used to retrieve passages for answer generation. The evaluation covers HotpotQA, MuSiQue, and PopQA, with comparisons against five GraphRAG systems and a vector-based baseline. The authors also investigate routing components, retrieval budgets, corpus growth, and incremental index maintenance.

Critical strengths

S1: I found the grounding-oriented perspective compelling. Much of the GraphRAG literature concentrates on graph construction or evidence expansion, whereas this paper asks what happens when the initial entity selection is wrong. More importantly, the authors do not rely exclusively on comparisons with complete external systems. In Table 3, replacing the learned router with vector retrieval while retaining the same fine-tuned encoder decreases EM on the three datasts. These comparisons provide more convincing evidence for the router than the main leaderboard results. The multi-parent and path-accumulation ablations further support the proposed design.

S2: Avoiding API-based LLM judgments during grounding is a useful design choice. On HotpotQA, DAG-RAG uses less seconds and tkens than for KAG. I would not interpret this as an unconditional efficiency advantage, since simpler retrieval approaches remain faster, but the quality-cost trade-off against more elaborate GraphRAG pipelines is promising.

Critical weaknesses

W1: The improvements in Table 1 cannot be attributed entirely to hierarchical grounding (if one looks at Appendix A.1 and Appendix E). The shared-encoder ablation partially addresses this concern, and I consider it one of the strongest parts of the evaluation. However, it remains unclear how much of the advantage would persist against better-adapted flat or hierarchical retrievers under equivalent supervision. There are also metric-dependent differences worth acknowledging. For instance, KAG achieves higher F1 on MuSiQue despite DAG-RAG having higher EM. The result is encouraging, but less uniformly favorable than the broad performance claims might suggest.

W2: The evaluation uses 250 questions per benchmark, with approximately 9–11k source passages per full corpus. The corpus-growth experiment is useful, but its largest setting is still relatively modest for a method motivated by large-scale GraphRAG. Finally, offline construction and training costs are not reported.These omissions matter because the paper positions DAG-RAG as both scalable and adaptable. The results establish a promising online retrieval mechanism, but not yet the full operational advantage of maintaining a trained hierarchical index as knowledge changes.

Questions for authors

Q1: Could the authors provide a more tightly controlled comparison against flat and hierarchical retrievers trained with the same oracle entity supervision, encoder, and comparable computational budgets? I would particularly like to know whether path conditioning remains beneficial when the alternatives receive equivalent training.

Q2: How stable is grounding quality after incremental updates, particularly when new entities alter cluster boundaries or introduce additional parent links?

Q3: How sensitive is the router to noise in the oracle seed entities used for training?

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

Related papers

Discussion (0)

Sign in to comment.