acceptodds
Under review as a conference paper at ICLR 2027

Towards a Transferable Foundation Retriever for Multi-Hop Reasoning over Knowledge Graphs

Abstract

Foundation retrievers aim to reuse retrieval knowledge across knowledge graphs (KGs), reducing the need to train a separate model for each graph. For multi-hop reasoning, this requires identifying relevant relations and ranking answers under changes in graph structure and question distribution. Yet strong in-domain performance does not necessarily translate into effective cross-KG retrieval. Our analysis of KGFR reveals a substantial ranking gap: correct answers often remain reachable but receive low ranks after transfer. We propose KGSeek, a multi-source foundation retriever that uses semantic relevance to guide retrieval and graph structure to distinguish candidate answers. KGSeek learns question–relation relevance from source questions and answers. It aggregates this relevance along paths to form a semantic ranking reference. A shared scorer incorporates neighborhood context through bounded adjustments. Length-normalized path scoring preserves this bound at the entity level, allowing contextual refinement while retaining strong semantic preferences. Joint training across source KGs learns shared scoring functions that apply to new graphs through their relation labels and structure. Experiments on Freebase and DBpedia show that one source-trained model improves cross-KG retrieval and downstream question answering (QA) over strong baselines with different large language model (LLM) readers. The same architecture also improves in-domain retrieval and maintains competitive QA performance. These results show the value of semantic guidance for transferable multi-hop retrieval.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Knowledge Graphs, Multi-hop Question Answering, Foundation Retriever

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Kp6cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies foundation retrievers for multi-hop question answering over knowledge graphs. Its goal is to train a retriever on one or more source KGs and transfer it to a different KG, where relation labels, graph structure, and question distributions may change. The authors analyze a prior KG foundation retriever and report that correct answers often remain reachable after transfer but receive poor ranks. They propose KGSeek, which combines question–relation semantic relevance with graph-neighborhood structure, aggregates relevance along paths, and applies bounded, length-normalized contextual adjustments. Joint training over source KGs is intended to produce scoring functions that transfer through relation labels and graph structure. Experiments on Freebase and DBpedia report improvements in cross-KG retrieval and downstream QA, as well as competitive or improved in-domain results with different LLM readers. The main unresolved issue for me is whether this transfer setting has a compelling practical role relative to adapting or fine-tuning a retriever on the target KG using target-graph training data.

Critical strengths
  1. The paper identifies a concrete ranking failure mode in cross-KG transfer: correct answers may remain reachable but receive low ranks because the retriever does not transfer its source-KG preferences effectively. Combining semantic relevance with graph structure is a reasonable response to that failure mode.

  2. The evaluation appears to consider both retrieval and downstream QA, and it reports results across Freebase and DBpedia with multiple LLM readers. This is useful evidence that the proposed scoring architecture is not limited to a single reader or a single in-domain benchmark.

Critical weaknesses
  1. The practical use case for zero-shot or source-trained transfer is not established. In a practical KG-based QA system, the retriever is often adapted or fine-tuned to the target graph when target-graph questions, answers, or retrieval supervision are available. Such adaptation may be preferable because it can directly exploit the target graph’s schema, relation semantics, and question distribution. I do not assume that target supervision is always available: transfer could be valuable for newly created KGs, low-resource domains, or systems that must support many graphs. However, the paper does not identify and validate such a deployment regime, nor does it compare against target-KG adaptation and the associated quality/compute/data trade-offs.

  2. The paper’s main contribution may improve a benchmark transfer setting without showing a consequential downstream benefit. Cross-KG retrieval is an interesting scientific problem, but the experiments do not demonstrate that KGSeek reduces deployment cost, eliminates target-KG fine-tuning, improves QA in a realistic low-resource regime, or enables a use case that target-specific training cannot support. The reported cross-KG gains therefore do not yet establish significance for a broad conference. A stronger paper would show when transfer is preferable to adaptation, how much target data it saves, and whether its advantage persists under matched compute and retrieval budgets.

Questions for authors
  1. Who is the intended user or deployment scenario for a source-trained retriever on a new KG? Please describe a realistic setting in which target-KG questions/answers or retrieval labels are unavailable, delayed, or too expensive to obtain, and explain why KGSeek is preferable to an unsupervised retriever in that setting.

  2. How does KGSeek compare with fine-tuning or otherwise adapting the same retriever on the target KG when a modest amount of target-graph supervision is available? Please report quality-versus-target-data and quality-versus-compute trade-offs, including the point at which target adaptation overtakes zero-shot transfer. This comparison is necessary to establish when the proposed transfer setting is practically preferable rather than merely possible.

  3. Do the cross-KG gains remain when the target KG has relation labels or graph schemas that are not aligned with the source KGs? The method appears to rely on relation labels and shared scoring functions; please clarify what information is assumed to be shared or available at transfer time.

  4. Are the downstream QA gains driven by improved retrieval itself, or by differences in retrieved context size, answer candidate set, or LLM reader? Please include matched-budget comparisons and report whether the gain persists against a target-specific baseline.

  5. What practical resource does the proposed method save—target-graph annotation, retriever training, inference cost, or engineering effort—and how is that saving measured? Without this analysis, it is difficult to assess the significance of the transfer objective.

Reviewer_M1H7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces KGSEEK, a multi-source foundation retriever for multi-hop reasoning over knowledge graphs that aims to transfer retrieval ability across KGs with different schemas and question distributions. The key idea is to learn a semantic reference ranking from source QA supervision using question–relation relevance derived from frozen text embeddings, and to refine it via a bounded, context-aware graph scorer. Trained on KQA Pro and MetaQA, KGSEEK improves cross-KG retrieval (Hits@1/MRR) and downstream QA F1 on WebQSP, CWQ, and LC-QuAD compared to the baselines (KGFR, GFM-RAG, G-reasoner).

Critical strengths

S1. The paper offers a novel perspective on transferability in cross-KG retrieval.

S2. The experimental results demonstrate the effectiveness of the proposed method.

Critical weaknesses

W1. The proposed approach appears to have limited practical applicability. Real-world applications operate in open environments spanning many domains. The paper’s approach seems to require learning a broadly transferable retriever similar to a pretrained model, yet its evaluation includes only two source datasets and three target datasets. This limited coverage does not adequately reflect the diversity and complexity of real-world settings.

W2. The experiments have limited dataset coverage and insufficiently baselines. Given the paper’s emphasis on cross-KG transfer, evaluation on sufficiently diverse graphs and different retrieval approaches is needed to establish its applicability and relative advantages. However, training uses only KQA Pro and MetaQA, while the three target datasets cover just two KG families, Freebase and DBpedia. Comparisons are also limited to three graph-based retrievers (KGFR, GFM-RAG, and G-reasoner) with the latter two adapted for entity retrieval. These experiments support performance gains under the evaluated settings but do not sufficiently establish generalization across diverse domains, graph schemas, and question types. They also limit conclusions about the method’s advantages over other retrieval approaches.

Questions for authors

See the weaknesses.

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

The paper proposes KGSeek, a retriever for multi-hop KGQA meant to transfer to new knowledge graphs without target training. It scores each edge by question–relation cosine similarity plus a learned, bounded correction, refines this with neighborhood context from a Graph Transformer, and ranks entities by their best path. It is trained on KQA Pro and MetaQA and tested on WebQSP, CWQ (Freebase) and LC-QuAD (DBpedia). It reports large cross-KG gains over KGFR; for example, WebQSP H@1 rises from 9.28 to 37.10.

Critical strengths

The cross-KG retrieval gains are large and consistent across three datasets and two target KGs, with paired bootstrap intervals (Table 9). The analysis in Figure 1, showing that answers remain reachable but are ranked poorly after transfer, is a useful motivation.

Critical weaknesses
  1. The paper’s key idea is that graph context refines the semantic ranking. However, Table 16 shows that in cross-KG evaluation, the learned reference S0S_0 (without context) beats the final score SθS_\theta on both WebQSP (H@1 48.28 vs. 37.10) and CWQ (17.33 vs. 11.92). Even frozen cosine similarity beats SθS_\theta on CWQ H@1 (14.10 vs. 11.92). The paper still deploys SθS_\theta and reports it as the main result. This undermines the claimed role of bounded context, and Proposition 1 offers little practical value if context makes rankings worse.

  2. Appendix C states that WebQSP/CWQ validation and “earlier test analyses informed method development”. Two of the three “unseen” targets were therefore used during development and only LC-QuAD is truly held out. In addition, all results come from a single training seed, and the bootstrap intervals cover only question sampling, not training variance.

  3. GFM-RAG and G-reasoner are adapted by the authors from document retrieval, and their cross-KG scores are close to zero (e.g., CWQ H@1 of 1.64 and 2.18). This suggests the adaptations may not be well tuned. Closer KGQA retrievers, such as SubgraphRAG (Li et al., 2025) and Yao et al. (2025), are cited but not compared.

  4. Only about 1,364 training questions are used per source, and MetaQA is templated. This is a narrow basis for a “foundation” retriever.

  5. In Proposition 1, with B = 0.3 and K = 1, the bound only protects rankings with cosine gaps above 0.6. Such gaps are likely rare, so the guarantee may be vacuous in practice.

Questions for authors
  1. Why do you deploy SθS_\theta when S0S_0 performs better in the cross-KG setting (Table 16)? What does contextual scoring contribute at inference?
  2. How exactly did WebQSP/CWQ test analyses inform method development? Can you report results with these targets fully held out?
  3. Can you report results over multiple training seeds?
  4. How were GFM-RAG and G-reasoner tuned after adaptation, and can you compare with SubgraphRAG or Yao et al. (2025)?
  5. How often do candidate pairs in practice have score gaps larger than 2B_K, so that Proposition 1 applies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.