acceptodds
Under review as a conference paper at ICLR 2027

Simple Links, Strong Retrieval: Local Entity Expansion for Multi-Hop Evidence Retrieval

Abstract

Retrieving complete supporting evidence from document collections remains a central challenge for knowledge-intensive language systems, particularly when answering a question requires connecting information across multiple sources. Similarity-based retrieval can overlook passages whose relevance depends on intermediate evidence, while graph-based and iterative approaches often introduce query decomposition, global propagation, or repeated reasoning to recover these connections. We present SCOPE, an entity-guided retrieval framework that addresses this challenge through a deliberately simple design: local evidence expansion followed by explicit evidence selection. SCOPE turns precomputed passage–entity links into access paths beyond the initial dense ranking, then combines semantic, lexical, and entity-based signals with listwise reranking to select supporting passages. This modular design enables complementary evidence discovery without query decomposition or global graph propagation. On the established multi-hop QA benchmarks MuSiQue, 2Wiki, and HotpotQA, SCOPE outperforms leading graph-based baselines, including BrowseNet and HippoRAG 2, in both evidence retrieval and downstream question answering. On LoCoMo, however, reranked graph-free retrieval performs better, highlighting a limitation of entity-guided expansion in conversational memory. These results demonstrate that a simple local-expansion design can deliver stronger retrieval and answer quality than more elaborate retrieval pipelines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Retrieval-Augmented Generation, Evidence Retrieval, Multi-Hop Question Answering, Entity-Guided Retrieval, Graph-Based 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_hwZHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SCOPE, a multi-hop retrieval method based on entity-guided passage expansion. It first retrieves passages using dense retrieval, expands the candidates through shared entities, and then uses rank fusion and LLM-based selection to retrieve supporting evidence.

Critical strengths
  1. The problem of multi-hop evidence retrieval is important for RAG.

  2. The paper provides detailed ablation studies and evaluates different embedding models.

Critical weaknesses
  1. The novelty is very limited. The main idea is to expand retrieved passages through shared entities, followed by rank fusion and LLM-based reranking. These are well-established techniques in RAG and GraphRAG. In particular, KG²RAG [1] already uses initially retrieved passages as seeds and performs knowledge graph-guided chunk expansion. Other related works [2] have also explored graph-based evidence retrieval and multi-hop retrieval. I do not see a clear technical contribution beyond combining existing components.

  2. The improvements over HippoRAG 2 are not convincing. In Appendix C.4 (Table 23), when HippoRAG 2 is equipped with the same LLM-based evidence selector, SCOPE does not achieve statistically significant improvements in retrieval recall or QA F1 across the four tested settings. This suggests that the improvements in the main tables may largely come from the LLM-based selection rather than the proposed entity expansion.

  3. The evaluation is limited to traditional benchmarks. Most experiments use MuSiQue, 2Wiki, and HotpotQA, which have been widely used in previous studies. On LoCoMo, the proposed method performs worse than graph-free hybrid retrieval with selection. It is unclear whether the approach can generalize to other domains and retrieval tasks.

  4. The efficiency advantage is unclear. Although the paper emphasizes the simplicity of SCOPE, it still requires LLM-based indexing and evidence selection. According to Table 7, SCOPE uses more indexing and online LLM tokens than HippoRAG 2 on 2Wiki. Therefore, the proposed method does not demonstrate a clear efficiency advantage.

[1] Zhu et al. Knowledge Graph-Guided Retrieval Augmented Generation (KG²RAG).

[2] Asai et al. Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering.

Questions for authors

Please refer to the weaknesses.

1 comment

Thank you for the detailed review.

W1 — novelty vs KG²RAG and Asai et al. KG²RAG expands seed chunks through an LLM-extracted triple KG and organizes a subgraph as context. Asai et al. learn a recurrent retriever over Wikipedia hyperlinks. SCOPE uses a passage–entity bipartite index built from entity mentions, takes a single hop, and does no graph-side reasoning or learning at query time. We agree the components are known. The paper's contribution is the controlled dissection (see our reply to WDWC W1). In the revision we will cite both works in §2 and state the differences.

W2 — are the gains from the selector rather than the expansion? Table 15 was designed to answer this question; see the general response. With the selector held fixed, removing the expansion lowers R@10 significantly in 10/10 (corpus × selector) cells on 2Wiki and HotpotQA. On HotpotQA the loss is about the same whichever selector is used (+4.4 to +5.1 pp). On MuSiQue the effect is smaller: positive in 9/10 cells, significant under the two LLM selectors in R@10 and in 7/10 cells in F1 (Table 15d). So the selector does not create the candidate gain; it determines how much of it reaches the reader. We agree with the second part of your reading: with the same selector, SCOPE does not beat HippoRAG 2. Pooled over the four Table-23 settings plus HotpotQA, the two are equivalent within ±1 pp on R@10 and F1, and HippoRAG 2 is better at R@5. We will remove the "outperforms HippoRAG 2" claim in the revision.

W3 — benchmarks, LoCoMo, generalization. See our reply to TbzC W3. In the revision we will move the LoCoMo analysis (entity sparsity, within-session coreference) from the appendix into §6, framed as a failure condition of entity-based expansion, together with the comparison-question result above.

W4 — efficiency. We agree that Table 7 does not support an indexing-cost advantage, and we will remove that wording. The data support an online-cost advantage. Under a shared selector and equal quality, SCOPE uses 4,175 online tokens per HotpotQA question against 7,132 for HippoRAG 2 + selector (Table 15a). Measured sequentially on 200 HotpotQA questions, it answers in about 2.7 s per question against about 3.9 s. Selector and reader times are identical, and the gap comes from HippoRAG 2's query-time retrieval (Table 15e). The revised efficiency paragraph will read "comparable indexing cost, lower online cost and latency".

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

The paper proposes SCOPE, a multi-hop retrieval pipeline. It expands the top-5 dense passages by one hop through precomputed passage–entity links, fuses dense, BM25 and entity-overlap rankings with RRF, and uses a gpt-4o-mini selector to choose the final evidence set. With a MiniLM backbone, SCOPE reports higher recall and QA scores than HippoRAG 2 and BrowseNet on MuSiQue, 2Wiki and HotpotQA.

Critical strengths

S1: The method is simple and modular. It only performs a one-hop expansion over passage–entity links, without query decomposition or global graph propagation, so it is easy to implement and to plug into standard dense retrieval pipelines. S2: The ablation studies clearly quantify the contribution of each component.

Critical weaknesses

W1: The novelty is limited. Each component, including one-hop entity expansion, RRF and LLM-based selection, is standard. W2: The baselines are insufficient. IRCoT is discussed in the introduction but not compared. Widely used GraphRAG methods such as LightRAG [1] and RAPTOR [2] are also not evaluated.

[1] LightRAG: Simple and Fast Retrieval-Augmented Generation. [2] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval.

Questions for authors

Refer to the Weaknesses.

2 comments

Thank you for the review.

W1 — novelty. We do not claim a new operator. One-hop entity expansion, RRF and LLM selection are each standard, and we say so. The contribution is a controlled account of when a one-hop passage–entity expansion helps, and how much of the end-to-end gain each stage is responsible for. This rests on candidate-count-matched controls (Tables 3/10), the shared-selector comparison with HippoRAG 2 (Tables 23/15a/15e), the gold-connectivity predictor (Table 6), the distractor-scaling study (Table 13), and the new selector-swap study (Table 15). One outcome of this dissection is the interaction in the general response: expansion adds candidates regardless of the selector, but only a strong set-level selector turns them into QA gains. In the revision the contribution list will say exactly this.

W2 — IRCoT, LightRAG, RAPTOR. IRCoT is an iterative retrieve-and-reason loop with several LLM calls per question, while SCOPE is single-pass. In the revision we will present IRCoT as motivation, not as a comparable system. For RAPTOR and LightRAG we rely on HippoRAG 2's evaluation on the same three corpora (Gutiérrez et al., arXiv 2502.14802, Tables 2–3). There, RAPTOR reaches passage Recall@5 of 57.8 / 66.2 / 86.9 (MuSiQue / 2Wiki / HotpotQA) against HippoRAG 2's 74.7 / 90.4 / 96.3. LightRAG produces no passage-level ranking and reaches 1.6 / 11.6 / 2.4 QA F1, against 48.6 / 71.0 / 75.5 for HippoRAG 2. Under a shared selector SCOPE is equivalent to HippoRAG 2 within ±1 pp (R@10 and F1, pooled over 4,544 questions), so a direct run of these two systems is very unlikely to change their ranking relative to SCOPE. We will add this reasoning and the citations to §5.

Thanks for the response. I will keep my score.

Reviewer_TbzCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents SCOPE, a multi-hop evidence retrieval framework that expands dense retrieval results through a single passage–entity–passage step and applies an LLM-based evidence-set selector for evidence selection. Experiments show improvements over the evaluated baselines on multi-hop QA benchmarks but weaker performance than graph-free hybrid retrieval with selection on conversational memory.

Critical strengths
  • The paper presents a simple and interpretable retrieval framework that uses local entity expansion to recover evidence missed by dense retrieval.
  • Extensive ablations and candidate-count-matched controls provide useful insights into the contributions of candidate expansion, fusion, and evidence selection.
Critical weaknesses
  1. Limited comparisons with alternative methods. While this paper shows the importance of applying the LLM-based selector, it does not compare with established cross-encoder or alternative LLM-based rerankers with the same LLMs for same model sizes. Also the paper does not include comparisons with recent agentic search approaches like NEMO or DCI. Therefore, the experiments do not establish the advantage over existing reranking approaches.
  2. Insufficient evidence for the advantage of the structured representation. Although the paper compares against graph-based retrieval methods, the results do not convincingly establish the advantage of its proposed text organization. For example, Table 23 shows no statistically significant advantage over HippoRAG2 when both methods use the same selector. Broader comparisons with recent memory organization methods would further strengthen the evaluation.
  3. Limited evaluation on long-form narrative understanding, which is usually used for evaluation of structured representation approaches.
Questions for authors

Please refer to the missing baselines in my review.

2 comments

Thank you for the careful review.

W1a — cross-encoder and same-LLM rerankers. We ran this comparison (general response; to be added as Table 15b). On identical candidate pools, R@10 for fusion / MiniLM-CE / bge-reranker-v2-m3 / gpt-4o-mini pointwise / gpt-4o-mini listwise is: MuSiQue 0.625 / 0.650 / 0.692 / 0.632 / 0.703 (F1 0.334 / 0.357 / 0.441 / 0.353 / 0.463); 2Wiki 0.803 / 0.782 / 0.814 / 0.811 / 0.891; HotpotQA 0.890 / 0.911 / 0.929 / 0.909 / 0.930. bge-reranker-v2-m3 is a credible zero-API alternative on two of the three corpora, and we report it as such. Pointwise use of the same LLM does not beat the fusion order, so the listwise set-level judgement is what matters.

W1b — agentic search (NEMO, DCI). SCOPE is a single-pass retriever: one dense call, one precomputed one-hop expansion, one selector call, about 4.2k online tokens per question. Agentic search methods are iterative outer loops that can call any retriever, SCOPE included, as a tool. A head-to-head comparison would conflate the retriever with the search policy and with a much larger token budget. In the revision we will discuss NEMO and DCI in Related Work and state the single-pass setting explicitly.

W2 — advantage of the structured representation; Table 23. We agree that the cross-system comparison does not establish an advantage, and we will revise the claim accordingly (general response). Pooled over all five shared-selector settings, SCOPE and HippoRAG 2 + selector are equivalent within ±1 pp on R@10 and F1, and HippoRAG 2 is better at R@5. What the one-hop expansion contributes shows up in the within-system, same-selector control (Tables 15c/15d): removing it lowers R@10 on 2Wiki and HotpotQA under every selector. On MuSiQue the drop is smaller and significant only under the LLM selectors. The by-type breakdown shows where the benefit comes from. On HotpotQA under the shared selector, SCOPE reaches R@10 0.939 vs 0.930 on bridge questions and 0.916 vs 0.971 on comparison questions. The expansion helps when a bridge entity exists and is neutral to harmful otherwise, consistent with §6 and Table 13e. The argument for SCOPE is therefore equal accuracy with a simpler and cheaper mechanism, not higher accuracy.

W3 — long-form narrative understanding. We agree this is a limitation. LoCoMo is the one long-context setting we report, and SCOPE loses there to graph-free hybrid retrieval (Table 10c), which we analyse rather than omit. Entities are sparse and coreference within a session is frequent, so one-hop entity links carry little signal. Building SCOPE and HippoRAG 2 indices for a narrative benchmark such as NarrativeQA was not feasible within the discussion period. We will list it in Limitations as the main open evaluation.

Regarding the response to W1, I agree with the authors that NEMO and DCI incur higher costs. However, many commonly used LLM rerankers are both listwise and cross-encoder, which should be compared. For example, given that single-step NEMO consists of embedding retrieval followed by an LLM reranker, NEMO’s reranker would provide a strong single-step baseline.

Moreover, there are published listwise LLM rerankers, such as GroupRank and other leading rerankers on the BRIGHT leaderboard. QRRanker also reports results on HotpotQA, LoCoMo, and NarrativeQA.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.