acceptodds
Withdrawn from ICLR 2027

LLM-KG Collaborative Reasoning with Knowledge Probes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Language Models (LLMs) excel at reasoning but suffer from static knowledge and hallucinations in fact-intensive tasks. Knowledge Graphs (KGs) offer trustworthy evidence, yet existing KG-centric agents remain brittle: relevant evidence may be structurally unavailable due to missing entities or relations, or computationally unavailable when buried in dense neighborhoods that are too costly to traverse exhaustively. We propose a KG-first collaborative reasoning framework that improves access to existing KG evidence in dense neighborhoods while recovering from missing KG links without relinquishing factual control to the LLM. We introduce the Guidance Graph of Thought (GGoT) as the framework's state interface, exposing unresolved reasoning slots as probeable states, and build on this interface with a conditional KG-first protocol. This protocol preserves reasoning continuity when required KG links are absent, reduces unnecessary exploration when relevant facts exist in large neighborhoods, and admits LLM-generated knowledge only under strict factual control. Experiments on Freebase- and Wikidata-based QA benchmarks show consistent improvements over strong KG-LLM reasoning baselines across both large and smaller LLMs. Efficiency and ablation analyses further confirm that the framework reduces dense-neighborhood exploration while making controlled recovery from KG dead ends beneficial rather than permissive.

Primary area:
interpretability and explainable AI
Keywords:
knowledge graph, question answering, large language model

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

This work aims to improve LLM-based KGQA performance in two ways: (1) when the essential triple for reasoning is structurally missing from the KG; (2) when the triples are computationally buried in dense neighbourhoods. To address this issue, it first parses a question into triples with specific and generic entities. The LLM may propose candidate "probes" (i.e., instantiations) for triples with exactly one specific end. In dense neighbourhoods, probes are intersected with all neighbours to get a relation context. When local exploration fails, probes may create existence-grounded or virtual bridges, which are kept only if later reasoning stays consistent.

Critical strengths

The example in Figure 3 is clear and easy to follow.

Critical weaknesses
  1. The answer-generation prompt (Table 13) instructs the model to "answer with your own knowledge" when the triples are insufficient, and Section 4.1 describes an LLM-only fallback when no anchor is found, which contradicts the claim of not "relinquishing factual control to the LLM". -- this is a soundness issue.

  2. The claim of recovering from structurally missing KG links is never tested in a controlled setting, since all the experiments use the standard WebQSP/CWQ splits over full Freebase or Wikidata, and the paper itself states that its case study "does not establish missing-edge repair" -- so the claim is not substantiated.

  3. Regarding the first contribution, the term "interface" is not clearly defined, making it difficult to understand what the proposed contribution actually entails. Overall, the technical novelty appears limited, and the work seems to be largely an incremental engineering effort rather than a substantive research contribution.

Questions for authors

In Table 1, why are the results for CWQ-partial and CWQ-complete all the same?

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

This paper proposes a KG-first collaborative reasoning framework that uses LLM-generated knowledge probes to guide graph exploration. A Guidance Graph of Thought represents specific entities and unresolved entity slots. Probes help narrow candidate neighborhoods when the KG is dense and propose tentative bridges when local exploration fails, while consistency checks determine which continuations are retained. Experiments on Freebase- and Wikidata-based QA benchmarks report improved answer accuracy over the evaluated baselines, supported by component ablations and computational cost analyses.

Critical strengths
  • The paper identifies two practical obstacles to KG-based reasoning: missing connections and costly exploration of dense neighborhoods. Using generated candidates as search guidance, with separate checks before commitment, provides a clear motivation for the framework.

  • The experiments evaluate multiple backbones and KG settings. Component ablations and answer-level trust-filtering comparisons provide evidence that probing and filtering contribute to downstream QA performance.

Critical weaknesses
  • The scope of the probing mechanism remains unclear. Section 3.2 restricts probes to entity completion in Type II triples, treating relations as predicate constraints. How does the framework handle queries where the entities are known but the connecting relation or path structure must be discovered?

  • The factual reliability of accepted bridges is insufficiently evaluated. What operational criteria determine when a provisional bridge becomes trusted? Please report the correctness of newly accepted links in addition to downstream answer accuracy.

  • The benefit of KG-local acceleration is not sufficiently isolated. How much search and computation does neighborhood narrowing save in dense neighborhoods, and how much relevant-candidate coverage does it lose?

  • The evidence for missing-link recovery remains limited. The qualitative examples mainly demonstrate recovery of existing KG evidence despite noisy probes, rather than missing-link repair. Could the authors systematically remove relations from known answer paths and evaluate answer recovery under different levels of KG incompleteness across backbones?

Questions for authors

Please refer to Weaknesses.

Reviewer_td1pOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes an LLM–KG collaborative reasoning framework. The key idea is to leverage the parametric knowledge of LLMs to improve reasoning over knowledge graphs. The framework introduces the Guidance Graph of Thought (GGoT) to represent reasoning states and employs LLM's knowledge probes to assist KG exploration. Specifically, the proposed method uses knowledge probes to identify relevant relations in dense KG neighborhoods and introduces trust-aware bridging to recover reasoning paths when required KG connections are missing. The authors evaluate the proposed method on multiple benchmarks and the experimental results demonstrate improvements over some training-free LLM-reasoning baselines.

Critical strengths
  • The paper is well written.
  • This paper provides comprehensive analysis and ablation studies.
Critical weaknesses
  • The novelty is limited. The core idea of leveraging LLM parametric knowledge to complement incomplete or insufficient KG evidence has already been explored in several prior studies such as GoG [1] and PARoG [2].
  • The proposed framework is training-free and explicitly relies on the parametric knowledge of the underlying LLM to generate useful knowledge probes. Therefore, its effectiveness is likely to depend substantially on the knowledge and reasoning capabilities of the backbone model. However, the main experiments primarily rely on outdated open-sourced LLMs. This paper does not evaluate the proposed method with recent closed-source LLMs, such as the GPT-5 series or recent Claude and Gemini models. For a training-free framework intended to be broadly applicable across different LLMs, evaluating stronger and more diverse backbone models is important. In particular, it remains unclear whether the proposed knowledge probing mechanism continues to provide substantial improvements when the underlying LLM already possesses stronger factual knowledge and reasoning capabilities.
  • Although the proposed method improves over some baselines, the performance is not competitive compared with recent SOTA methods such as PARoG and SoG [3].

[1] Generate-on-Graph: Treat LLM as both Agent and KG in Incomplete Knowledge Graph Question Answering. Xu et al. EMNLP 2024.

[2] Plan-Answer-Refine-on-Graph: Structured Planning and Self-Refinement for Large Language Model Reasoning on Knowledge Graphs. Shi et al. ICLR 2026.

[3] Search-on-Graph: Iterative Informed Navigation for Large Language Model Reasoning on Knowledge Graphs. Sun et al. KDD 2026.

Questions for authors

Please address the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.