acceptodds
Under review as a conference paper at ICLR 2027

ReScopeRAG: Evidence-Guided Retrieval Scope Adaptation for Multi-Step Reasoning

Abstract

Retrieval-augmented generation (RAG) relies on effective retrieval to identify relevant evidence from large and complex corpora. Existing approaches improve evidence acquisition through diverse retrieval strategies, but typically fix the retrieval scope from the initial query. This limitation is particularly pronounced in multi-step reasoning, where intermediate evidence can reveal new targets and evolve the information need beyond the initial scope. We propose ReScopeRAG, an iterative graph retrieval framework for multi-step reasoning. It dynamically adapts the retrieval scope as accumulated evidence reveals new targets. The framework first models dependencies among sub-queries to organize the reasoning process. A Master–Sub-Agent architecture then coordinates evidence acquisition and scope adaptation, enabling the system to progressively shift the retrieval scope as new targets emerge. Experiments on multiple general multi-hop question answering datasets show that ReScopeRAG achieves superior performance compared with conventional retrieval, graph-based, and agentic RAG baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Retrieval-Augmented Generation, Multi-Hop Reasoning, Graph Retrieval, Adaptive 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_ptzrOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ReScopeRAG, a training-free graph RAG framework that splits a question into dependent sub-queries and uses a Master with several Sub-Agents to run PPR inside Leiden communities.

Critical strengths

Treating the retrieval scope as something that shifts during multi-hop reasoning is a reasonable idea, and the decomposition stats and case study show the placeholder chaining actually gets used.

Critical weaknesses

The QA gains are mixed: MA-RAG is far ahead on 2Wiki (about 13 F1), HippoRAG2 matches it on HotpotQA, and the text even says some agentic methods have higher average scores, which contradicts Table 1. The retrieval comparison isn’t fair, since a multi-round, LLM-filtered evidence pool is compared only against single-shot retrievers, with no iterative baselines like IRCoT; HotpotQA Recall@5 even drops well below HippoRAG2. There’s no cost analysis, even though each round involves several LLM calls (decomposition, filtering, navigation, sufficiency check, scheduling) across multiple Sub-Agents for up to six rounds. The ablations are hard to read: removing scope adaptation drops F1 below BM25, which looks like a strawman, the global-scope variant is below plain HippoRAG2, and community granularity, which the whole method depends on, is never studied.

Questions for authors

How sensitive are the results to the Leiden resolution, and what happens when the gold passages for one sub-query fall in communities that aren’t adjacent?

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

This paper proposes ReScopeRAG, an evidence-guided graph retrieval framework for multi-hop QA. It decomposes queries into dependent sub-queries and uses a Master–Sub-Agent architecture to iteratively retrieve evidence and adapt retrieval scopes. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue show competitive or superior average retrieval and QA performance. Ablation studies further verify the effectiveness of dependency-aware decomposition and retrieval-scope adaptation.

Critical strengths
  1. The paper identifies a clear multi-hop retrieval problem and provides a concrete mechanism for adapting retrieval scopes as new evidence emerges, with dependency-aware decomposition making evolving retrieval targets explicit.
  2. ReScopeRAG performs well, particularly on MuSiQue. Ablations consistently show the effectiveness of dependency-aware decomposition and scope adaptation.
Critical weaknesses
  1. The novelty of the proposed framework appears limited. Most components, including graph-based retrieval, query decomposition, and multi-agent coordination, build on existing techniques, while the paper does not clearly present how its retrieval-scope adaptation differs fundamentally from existing iterative or adaptive retrieval methods.
  2. The empirical validation is incomplete. Performance varies substantially across datasets, while uncertainty, computational-cost, and large-corpus scalability analyses are missing. Since larger or noisier retrieval settings are not tested, both the practical benefit and the claimed applicability to large and complex corpora remain unclear.
Questions for authors
  1. Why does ReScopeRAG achieve substantially lower F1 than MA-RAG on 2Wiki while performing strongly on MuSiQue? Could the authors clarify whether its benefits depend on particular dataset or reasoning characteristics?
  2. What are the average LLM calls, token usage, and latency per question for ReScopeRAG and the best baselines?
  3. The paper motivates ReScopeRAG for retrieval over large and complex corpora, but evaluates 1,000 sampled questions per dataset using their associated supporting and distractor passages. Can the authors report the resulting corpus/graph sizes and provide evidence that the method remains effective as corpus scale and distractor noise increase?
  4. Why are Sub-Agents with unresolved placeholders activated from the beginning? Does early activation improve retrieval performance compared with activating dependent agents only after their placeholders are resolved, and how much additional cost does it introduce?
Reviewer_whoJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses retrieval-augmented generation (RAG) for multi-step reasoning. The authors claim that the limitation of existing RAG methods is a static retrieval scope. To address this, they propose a method that clusters the entities of a pre-constructed knowledge graph into scopes and adaptively changes the scope while answering a question. The experiments include 12 baselines and 3 benchmarks.

Critical strengths
  1. The idea of changing the retrieval scope during multi-hop reasoning is intuitive.
Critical weaknesses
  1. The paper's premise is that existing retrieval methods are limited by a static retrieval scope, but this is only asserted in Figure 1 and illustrated with a single case study of the proposed method. No failure analysis of any existing baseline shows gold passages falling outside its retrieval scope. This is a strong, unverified assumption that the whole method is built on, and the authors have to justify it.
  2. The core mechanism is a known LLM-guided graph walk. An LLM ranks neighboring regions from an anchor entity, the chosen one is expanded, and the LLM judges sufficiency after each step. This is Think-on-Graph (Sun et al., ICLR 2024) with clusters as nodes and HippoRAG2 PPR as the local reader. The paper neither cites it nor compares against it.
  3. The proposed method relies heavily on the clustering of the graph. The clusters that define the retrieval scope are produced offline by Leiden with no relation to any question, so a scope may not contain what a question needs and the method may not work as expected. Experiments are needed that study the clustering under different parameters, such as cluster size or the clustering algorithm, to show the method is not an artifact of one particular partition.
Questions for authors
  1. Can you provide a failure analysis of HippoRAG2 showing what fraction of missed gold passages lie outside the initial retrieval scope, and whether the proposed method recovers specifically those cases?
  2. How does the method differ from Think-on-Graph, and can you report it as a baseline over the same graph with the same LLM?
  3. How sensitive are the results to the clustering? Please report performance under different cluster size caps and alternative clustering algorithms, and state the cluster size used in the main experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.