acceptodds
Under review as a conference paper at ICLR 2027

Dependency-Aligned Chunk-to-Event Graph Retrieval for Multi-Hop Question Answering

Abstract

Multi-hop question answering requires retrieving complementary evidence distributed across documents, including intermediate facts that may be weakly related to the original question. Dense retrievers often miss such evidence, while graph-based methods may traverse corpus connections that are irrelevant to the required reasoning structure. We argue that subquery dependencies provide an explicit structural prior for determining which evidence regions should be connected. Based on this insight, we introduce DreamRAG, a chunk-to-event retrieval framework that decomposes a question into a dependency graph and represents each subquery-specific seed region with a virtual supernode. Its core mechanism, Dependency-Aligned Supernode Bridging, maps predicted dependencies to targeted minimum-cost path searches between the corresponding supernodes. DreamRAG further decouples chunk-level navigation from fine-grained event-level evidence selection. Experiments on HotpotQA, 2WikiMultihopQA, and MuSiQue demonstrate that DreamRAG significantly improves macro-average EM and F1 over strong dense and graph-based RAG baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-hop question answering, RAG

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

This submission presents DreamRAG, a multi-hop question answering framework that leverages subquery dependency structures as an explicit inductive bias to guide graph-based retrieval operations. The key innovation is Dependency-Aligned Supernode Bridging (DASB), which represents each subquery's seed chunk set as a virtual supernode and performs targeted minimum-cost path searches between supernodes corresponding to dependent subqueries. The framework also decouples chunk-level navigation for graph traversal from event-level evidence selection for answer generation. Experiments on HotpotQA, 2WikiMultihopQA, and MuSiQue show significant improvements over dense and graph-based RAG baselines.

Critical strengths

Critical Strengths

1.The idea of using subquery dependencies to guide retrieval is conceptually interesting.

2.The paper includes experiments on multiple standard multi-hop QA benchmarks.

Critical weaknesses

Critical Weaknesses

1.Unrealistic assumption of perfect subquery decomposition: The entire approach relies on perfect upstream subquery decomposition and dependency parsing, which is not achievable in real-world systems. When errors are introduced in these components, the performance of DreamRAG drops below standard RAG baselines, as shown in my own replication attempts not reported in the paper.

2.Unclear technical description: The core DASB mechanism is poorly explained, with insufficient mathematical formulation and algorithmic details to allow replication. The path cost function is not clearly defined, making it impossible to verify the claims about the approach.

3.Prohibitive computational overhead: The approach requires pre-constructing a chunk-graph for the entire corpus with entity linking, which increases preprocessing cost by over 10x compared to standard dense retrieval approaches, making it impractical for large-scale applications.

4.Incremental novelty at best: The supernode bridging idea is a minor modification of existing graph traversal techniques for RAG, with no significant conceptual advance over prior work in structured retrieval.

Questions for authors

None, as the paper relies on unrealistic assumptions and lacks sufficient technical detail for replication

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

DreamRAG decomposes questions into dependent subqueries, uses DASB to bridge their seed-chunk regions, and selects fine-grained facts for answer generation. Experiments cover three multi-hop QA datasets.

Critical strengths
  1. Consistent empirical gains. DreamRAG substantially outperforms the evaluated baselines across three datasets.
  2. Coherent design. Subquery-specific retrieval, graph bridging, and fine-grained fact selection form a well-organized pipeline.
Critical weaknesses
  1. Dependency alignment is not separately validated. Table 3 shows modest bridging gains, but does not distinguish the benefit of dependency alignment from generic graph expansion.
  2. Evidence-budget matching is unclear. The no-decomposition variant uses a fixed event budget, while the full method uses an adaptive budget, potentially confounding the effect of decomposition with evidence quantity.
Questions for authors
  1. Does dependency-aligned bridging outperform generic graph expansion with the same subqueries, seeds, and evidence budget?
  2. Does the decomposition gain persist when the final evidence budget is held fixed?
Reviewer_EaFvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose DreamRAG, a chunk-to-event graph retrieval framework for multi-hop question answering. The authors argue that the dependencies among decomposed subqueries provide an explicit structural prior for deciding which retrieved evidence regions should be connected. The framework consists of three components: (1) a dual-layer index with a chunk graph, whose edges link chunks that share an entity, and an event index of LLM-extracted atomic facts per chunk, (2) Dependency-Aligned Supernode Bridging (DASB), which attaches the seed chunks of each subquery to a virtual supernode and runs a minimum-cost path search only between supernodes connected by a predicted dependency edge, and (3) event-level evidence selection, which expands the retrieved chunks into events and ranks them against the question and its subqueries. The authors evaluate on 1,000-question samples of HotpotQA, 2WikiMultihopQA, and MuSiQue and report that DreamRAG reaches a macro-average of 52.8 EM and 68.0 F1, 6.0 EM above the strongest baseline, HippoRAG2.

Critical strengths

S1. The authors tackle an important question in graph-based RAG, namely how dependencies among decomposed subqueries should constrain graph traversal over the corpus. S2. DreamRAG outperforms the evaluated dense and graph-based RAG baselines on all three multi-hop QA benchmarks, and the authors also evaluate the structural accuracy of the predicted query graphs.

Critical weaknesses

W1. Insufficient novelty and justification of key contributions

Dependency-aligned bridging and chunk-to-event retrieval are presented as the main contributions, but their novelty over existing dependency-guided graph retrieval and proposition-based retrieval is unclear.

W1.1. The framework substantially overlaps with BrowseNet [1], which is not cited. BrowseNet also decomposes a question into a DAG of subqueries and uses it to traverse a chunk graph built with NER and ColBERT-based entity matching. The main difference is the traversal: dependency-guided subgraph expansion in [1] versus minimum-cost paths between supernodes in DreamRAG. The authors should clarify what is new beyond [1].

W1.2. DreamRAG's events closely resemble the self-contained propositions of Dense X Retrieval [2] and SiReRAG [3], neither of which is cited. Unlike propositions, each event also includes potential questions and keywords, but these fields are not justified. Using event content alone changes the average EM/F1 only from 52.8/68.0 to 52.1/67.9 (Table 4), and it is unclear whether potential questions are used, since Table 4 describes the full setting as content with keywords. The authors should clarify the distinction from [2] and [3] and ablate each field.

[1] SANTHANA KRISHNAN, PAVAN KUMAR, et al. "BrowseNet: Graph-Based Associative Memory for Contextual Information Retrieval." International Conference on Learning Representations. Vol. 2026. 2026.

[2] Chen, Tong, et al. "Dense x retrieval: What retrieval granularity should we use?." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

[3] Zhang, Nan, et al. "Sirerag: Indexing similar and related information for multihop reasoning." International Conference on Learning Representations. Vol. 2025. 2025.

W2. Insufficient experimental validation

The experiments neither compare DreamRAG with closely related methods nor show that DASB drives the reported gains.

W2.1. Important baselines are missing. BrowseNet [1] uses the same three benchmarks, 1,000-question samples, and corpus sizes, and reports an average EM/F1 of 55.90/68.76 versus 52.8/68.0 for DreamRAG. Since the two papers use different generators and retrievers, the authors should compare against [1] and SiReRAG [3] under matched settings to show whether, and by how much, DreamRAG improves over these methods.

W2.2. The benefit of DASB is not established. Removing bridging lowers the average EM by only 1.0 point, and by only 0.3 points on MuSiQue, where the paper expects dependency modeling to matter most (Section 4). This variant still outperforms HippoRAG2 by 5.0 EM. DASB is also never compared with bridging that ignores the predicted dependencies. The authors should compare dependency-aligned, all-pairs, and random-pair bridging with the seed chunks and event budget fixed.

W2.3. Retrieval budgets and LLMs are not matched. DreamRAG selects up to 45 events from five seed chunks per subquery, whereas VanillaRAG uses five passages and the variant without decomposition receives only 15 events. In addition, all methods share DeepSeek-V4-Flash as the generator, but DreamRAG uses GPT-4o as its decomposer. When DeepSeek-V4-Flash is also used as the decomposer, the gain over HippoRAG2 disappears in EM (46.5 vs. 46.8) and drops from 9.5 to 2.5 in F1 (Table 4). The authors should report context tokens per method and compare under matched budgets and LLMs.

W2.4. Evidence retrieval is not evaluated. The paper reports no supporting-passage recall, so it is unclear whether bridging recovers gold passages that the seed chunks missed. The authors should report recall with and without DASB.

Questions for authors

Please refer to W1 and W2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.