acceptodds
Under review as a conference paper at ICLR 2027

From Query to Knowledge Topology: Agentic Knowledge Extraction for Multi-Perspective Graph Retrieval

Abstract

User queries can be partial and potentially ambiguous descriptions of the knowledge required to answer them. Retrieval based on a single query representation can therefore overlook relevant information, relationships, and alternative interpretations that remain implicit in the question. We propose a training-free framework for multi-perspective answer-set retrieval that reconstructs a candidate query-relevant knowledge topology through iterative LLM probing and uses this structure to guide retrieval from an external knowledge graph. Building on existing agentic knowledge-extraction methods, the framework explores the query through sequential associative probing, taxonomy exploration, and multi-perspective parallel probing. Extracted knowledge is filtered for relevance and redundancy and organized into an explicit structure of information, relationships, and perspectives. This structure provides a provisional model of the query’s knowledge requirements: its perspectives guide adaptive region retrieval, while its relationships inform the construction of supporting evidence through subgraphs. Retrieved entities seed a prize-collecting procedure over knowledge-graph relations, and the resulting subgraphs are organized into perspective-specific answer sets. The LLM then synthesizes these evidence sets into an answer that presents complementary findings and evidence-supported disagreements. We evaluate the framework on FB15K-237, WN18RR, PrimeKG, and NELL-995 using retrieval accuracy and answer-set distinctness.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
knowledge graphs, knowledge graph question answering, retrieval-augmented generation, subgraph retrieval, answer sets, multi-perspective retrieval, knowledge topology, large language models, neurosymbolic AI, representation learning, graph retrieval, agentic retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LpyWOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies multi-perspective retrieval from knowledge graphs for broad or ambiguous user queries. It proposes a training-free framework that first uses an LLM to reconstruct several query perspectives and taxonomy concepts, and then retrieves a separate evidence subgraph for each perspective.

The framework introduces an adaptive region retrieval method based on the local geometry of entity embeddings, followed by a prize-scored greedy subgraph construction procedure. Experimental results show that the reconstructed perspectives often collapse onto the same graph entities and that the proposed method generally does not outperform a simple nearest-entity Graph RAG baseline.

Critical strengths
  1. The paper reports its negative results honestly. Rather than presenting the framework as uniformly successful, the authors explicitly show that different perspectives often retrieve the same anchor entity and carefully analyze this failure through candidate-set overlap, embedding similarity, silhouette scores, and graph-structure statistics.

  2. The implementation and evaluation details are documented relatively thoroughly, including prompts, filtering thresholds, graph construction rules, calibration data, and additional structural measurements.

Critical weaknesses
  1. The methodological novelty is limited. The framework primarily combines LLM-based query expansion, cosine-similarity filtering, a Query2Box-inspired local region, and a heuristic greedy subgraph constructor. Each component is relatively simple, and the technical integration does not appear to introduce a sufficiently substantial algorithmic advance.

  2. The experimental results are not convincing, even though the baseline suite is relatively limited.

Questions for authors

Can the authors evaluate the method on a dataset with human-annotated ambiguous questions, perspectives and perspective-specific evidence?

How sensitive are the results to the region-construction and filtering hyperparameters? Please provide a sensitivity analysis to assess the robustness of the method to these choices.

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

This paper proposes a training-free framework for multi-perspective retrieval over knowledge graphs. It uses an LLM to generate query perspectives and associated parent–child concepts, filters these perspectives, retrieves candidate entities through adaptive embedding-space regions, and constructs perspective-specific evidence subgraphs for answer synthesis. Experiments on four knowledge graphs examine retrieval performance and the distinctness of the resulting evidence sets. The results are mixed: the framework improves over an unfiltered fixed-budget baseline, but does not consistently outperform simpler alternatives, and different perspectives frequently retrieve the same entities.

Critical strengths
  • The paper addresses a reasonable problem: a single query representation may overlook complementary aspects of an information need.
  • The framework explicitly separates model-generated hypotheses about information needs from evidence retrieved from an external knowledge graph.
Critical weaknesses
  1. The methodological contribution over existing work is unclear. Plan-on-Graph (PoG; Chen et al., 2024) already decomposes questions into sub-objectives to guide knowledge-graph exploration. Although expanding potential information needs differs conceptually from decomposing explicit query conditions, the paper does not sufficiently establish what its “topology reconstruction” adds beyond LLM-based query expansion or sub-question decomposition.

  2. The implementation does not fully instantiate the proposed formulation. The formal topology includes perspectives, hierarchical concepts, and candidate relations, but Appendix H explicitly states that the relation component is not constructed. The implemented reconstruction therefore consists primarily of LLM-generated perspective texts and parent–child concepts, rather than an explicit relational structure. This gap weakens the justification for presenting the method as knowledge-topology reconstruction.

  3. The most relevant comparisons are missing. The paper includes several retrieval controls, but does not compare against PoG or against ordinary query expansion and sub-question decomposition using the same downstream retriever. Such comparisons are necessary to establish whether the proposed perspective construction, filtering, and evidence organization provide benefits beyond simpler alternatives.

  4. The results do not adequately support the central multi-perspective retrieval claim. Different perspectives frequently retrieve the same entities, and the full framework does not consistently outperform simple neighborhood or raw-query retrieval baselines. Improvements over an unfiltered top-10 baseline are insufficient to establish the value of topology reconstruction, particularly when stronger controls remove much of that advantage. The final-answer checks are also lexical and do not establish improved correctness, coverage, or faithful presentation of disagreements.

  5. The explanation for the observed failure remains insufficiently isolated. The paper attributes limited evidence diversity largely to embedding geometry, but perspective generation, filtering thresholds, region construction, and graph coverage could also contribute. The negative results are informative, yet additional controlled experiments are needed to establish the proposed explanation.

Reference

Chen, L., Tong, P., Jin, Z., Sun, Y., Ye, J., and Xiong, H. (2024). Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs. Advances in Neural Information Processing Systems, 37, 37665–37691.

Questions for authors

See weaknesses above.

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

The paper proposes a training free framework for multi-perspective retrieval over knowledge graphs. When it comes to a vague question, a single query embedding could miss other possible information in the KG. The method asks an LLM to propose several perspectives of the query and retrieves a box-shaped region of the entities per perspective, and then grows a subgraph from those entities. An LLM then writes one answer from all the subgraphs. It is evaluated on 56 self-constructed questions per graph on FB15K-237, WN18RR, PrimeKG and NELL-995. The paper reports that graph RAG has higher F1 on nearly every family and graph, that the perspectives of a question mostly retrieve the same one or two entities, and concludes that the embedding geometry doesn't help separate perspectives.

Critical strengths
  1. Results are split by question family, paired bootstrap intervals and per-question win/loss counts are given, and the limitations section states the negative result honestly.
Critical weaknesses
  1. The method does not beat the baseline. Table 3 shows the subgraph beats the nearest-entity pile on 3, 17, 3 and 6 of 56 questions per graph, and loses on 27, 26, 25 and 26. The only wins are MRR on one-hop (all intervals include zero) and against top-10, which is weak. The paper’s own conclusion is that the framework’s value is “the hypothesis it isolates”.
  2. The main finding is already known since the text embeddings are anisotropic is cited from Ethayarajh (2019) and Mu & Viswanath (2018). The paper just remeasures it.
  3. The testing set is small and self-made with 56 questions per graph, 16 per path family, written by the same model family that proposes the perspectives.
Questions for authors
  1. On which task does the framework outperform the baseline? If none, what is the contribution of this work beyond the negative result in Section 5.3?
  2. Appendix sections D, F and G are empty. Were they meant to contain other aspects?
  3. The current comparisons use self-constructed questions and a reimplemented ToG. Please evaluate on a public benchmark (e.g., WebQSP/CWQ, PubMedQA/BioASQ) against standard RAG, GraphRAG (Edge et al., 2024), GIVE (He et al., ICLR 2025), and the original Think-on-Graph implementation (Sun et al., 2024), which are all training-free baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.