acceptodds
Under review as a conference paper at ICLR 2027

Structure-Internalized Rule Language Model for Faithful Knowledge Graph Reasoning

13 views

Abstract

Knowledge Graph Reasoning (KGR) aims to discover latent facts by leveraging the structural evidence available in KGs, posing a challenge to the structural semantic understanding capability of KGR models. Recent studies have demonstrated that Large Language Models (LLMs) can achieve remarkable progress on KGR tasks via flexible in-context learning. However, the inherent representation inconsistency between KG structural context and LLM parametric knowledge remains inadequately addressed. This limitation prevents LLMs from effectively perceiving reasoning evidence that aligns with KG constraints, which undermines both the effectiveness and faithfulness of reasoning. We refer to this problem as *reasoning evidence perception drift* of LLMs over KGs. To address this problem, we propose a Structure-Internalized Rule Language Model (SIRLM), which centers on structural rule generation to couple the parametric learning of structural knowledge with the faithfulness evaluation of reasoning logic, enabling LLMs to anchor tightly to KG-grounded evidence. Specifically, we first design a Structure-Internalized Rule Generator (SIRG), which incorporates an in-context learning block augmented with a structural relation memory to coordinate structural and parametric knowledge. Furthermore, we equip SIRG with a KG tokenizer based on structural invariance learning and a neuro-symbolic reasoner based on rule-constrained message propagation. These components provide SIRG with learnable structural representations and faithful rule-execution feedback, respectively. Our SIRLM can be seamlessly integrated into standard LLM training paradigms, such as SFT and GRPO. Extensive experiments against 17 state-of-the-art KGR methods on 36 datasets demonstrate the significant superiority of SIRLM.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Knowledge Graph Reasoning, Large Language Model, Rule Generation, Structure-Internalized Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Y6yeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SIRLM, a framework for knowledge graph reasoning that aims to mitigate reasoning evidence perception drift between LLM parametric knowledge and KG structural information. The method combines a structural KG tokenizer, a Structure-Internalized Rule Generator (SIRG) equipped with Structural Relation Memory (SRM), and a Rule-Constraint Message Propagation (RCMP) reasoner. SIRLM supports both supervised training and GRPO-based post-training. Experiments on 36 datasets demonstrate competitive performance across transductive and inductive KG reasoning settings.

Critical strengths

S1. Integration of structural knowledge and LLM-based reasoning.

The paper presents an interesting attempt to integrate structural KG representations with LLM-based rule generation. The SRM and RCMP modules provide a potentially useful approach to incorporating structural information into neural reasoning.

S2. Extensive evaluation across diverse reasoning settings.

The experiments cover 36 datasets, including challenging fully inductive settings, and compare against a broad range of existing approaches. The results demonstrate competitive performance, while the ablation studies provide evidence for the effectiveness of the proposed components.

Critical weaknesses

W1. Limited methodological novelty and unclear distinctions from existing approaches.

The proposed KG tokenizer largely builds upon existing structural representation learning techniques, such as those used in ULTRA and MOTIF, while LLM-based KG reasoning and rule generation have been explored in KRLM and ChatRule. The main contributions appear to be relation-memory-augmented attention and the incorporation of generated rule representations into NBFNet. However, the methodological distinctions and advantages over these existing approaches are not sufficiently established.

W2. Insufficient justification of the proposed rule-based reasoning mechanism.

The SRM module selects Top-K relations using learned relevance scores without explicitly establishing their logical relevance to the query. Moreover, RCMP incorporates the hidden states of generated rules into GNN initialization rather than explicitly enforcing the generated relation sequences during message propagation. Therefore, it remains unclear whether the improvements result from faithful rule-based reasoning or simply from additional structural representations.

Questions for authors

Questions For Authors

Q1. Novelty compared with KRLM.

KRLM already incorporates a structural KG tokenizer, a dynamic knowledge memory mechanism, and a structure-aware next-entity predictor. What are the fundamental methodological differences between SIRLM and KRLM, particularly regarding the proposed SRM and RCMP modules? Could the authors provide more direct evidence of their additional benefits?

Q2. Comparison with FtG.

FtG employs a structure-text adapter and ego-graph information to incorporate KG structural knowledge into LLMs. How does SIRLM's structural tokenization and rule generation provide advantages over FtG's structure-text fusion approach? Could the authors further clarify the differences and evaluate their respective effectiveness under comparable settings?

Q3. Necessity and faithfulness of rule generation.

Unlike FtG and KRLM, which predict target entities, SIRLM introduces an intermediate rule generation stage. However, RCMP incorporates rule hidden states into GNN message passing rather than explicitly executing the generated rules. What evidence demonstrates that this additional stage provides faithful reasoning rather than merely improving neural representations?

Q4. More targeted experimental comparisons.

Could the authors provide controlled comparisons with FtG and KRLM, including matched experimental settings and computational costs? In particular, ablations replacing SRM with KRLM-style knowledge memory or removing the rule generation stage would help establish the unique contributions of SIRLM.

Reviewer_96GROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates knowledge graph reasoning assisted by LLMs, integrating graph structure encoding, structural relation memory, and rule-conditioned message propagation. It employs supervised fine-tuning and reinforcement learning to bridge rule generation with entity prediction. The work encompasses transductive, entity-inductive, and fully inductive knowledge graph tasks, focusing on the interplay between structural information and LLM-based reasoning. However, the current manuscript lacks algorithmic-level constraints for its core definition of "faithful rule execution," and there are substantive inconsistencies between its training objectives and theoretical assumptions; consequently, the work is not yet sufficient to support its primary claims.

Critical strengths

By innovatively combining structured rule generation, relational memory, and neuro-symbolic reasoning, this approach enhances both the reliability and interpretability of reasoning. It has been extensively validated across 36 datasets and demonstrates strong generalization capabilities.

Critical weaknesses
  1. Equations (11)–(12) and Equation (20) in Appendix C perform message passing across all relations, with generation rules serving only for initialization. Since predictions are not constrained by rule paths or the order of relations, achieving a correct ranking does not equate to the faithful execution of a given rule.
  2. The outer negative sign in Equation (14) cancels out the minus sign preceding the negative sample term. Consequently, as the objective is minimized, a negative sample score approaching 1 would actually decrease the loss—a result that contradicts the goal of suppressing negative samples.
  3. Equation (15) constructs its objective by multiplying the sampled token probability by an advantage, rather than using a ratio of policy probabilities or log-probabilities. Its correspondence to the GRPO method mentioned in the text is unclear, and the actual optimization process cannot be determined from this equation.
  4. Neither Algorithm 1 nor Appendix H specifies whether the target edges (and their inverses) corresponding to the training queries are removed from the graphs used for structural encoding and rule mining. Thus, one cannot rule out the risk that the answer edges themselves are directly involved in constructing the supervisory evidence.
  5. In Appendix H, when no closed path is found, the query relation itself is used as the rule body, resulting in an identity rule (mapping a relation to itself). Such supervision fails to provide independent evidence for relational composition, yet it is used alongside multi-hop rules during training.
  6. Appendix D includes all memory vectors within subspace SS, which implies that the orthogonal drift is zero by definition. However, Assumption A posits that this drift decreases strictly as a function of the relation score; these two conditions cannot both hold true across varying score values.
  7. Assumption B2 in Appendix E presupposes that the argmax decoding operation is permutation-equivariant. However, structurally equivalent relations may yield identical scores, and the text does not define a mechanism for handling such ties while preserving equivariance—leaving a gap between the theoretical conclusions and the actual decoding process.
  8. The "average gain" reported in Table 1 is calculated by selecting the best result among four training paradigms for each metric and comparing it against the baseline average. This aggregation conflates different training conditions with reference benchmarks, and therefore does not accurately represent the gain of a specific configuration over the strongest comparable method. 9. Tables 1, 8, and 9 do not report the uncertainty associated with repeated training; therefore, the stability of the rankings of the results cannot be assessed, and the evidence for superiority across a large number of datasets fails to distinguish such performance from fluctuations caused by the randomness of the training process.
Questions for authors

The core issue is the lack of an effective link between rule-conditioned prediction and the faithful execution of those rules; furthermore, there is a directly verifiable discrepancy between the training loss and the underlying theoretical assumptions. While cross-scenario experiments hold some value, they are currently insufficient to bridge this gap in the method's effectiveness.

  1. How does SIRLM ensure faithful execution of generated rules when RCMP performs message passing over all relations rather than following the generated rule paths?
  2. Could the authors clarify the optimization objectives in Eqs. (14) and (15), particularly the negative-sample loss sign and the formulation of GRPO?
  3. Are target edges and their inverses excluded during structural encoding and rule mining to prevent information leakage? How are trivial identity rules handled during training?
  4. How are the reported average gains affected by differences in training paradigms and baseline selection? Could the authors provide run-to-run variance to establish the statistical reliability of the improvements?
Reviewer_LFt6OpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SIRLM, a structure-internalized rule language model aimed at resolving the tension between LLM parametric knowledge and structural evidence in KG reasoning. Specifically, the architecture brings together a structural-invariance KG tokenizer, a rule generator supported by structural relation memory, and a rule-constrained message passing scheme, utilizing both SFT and GRPO for training. The authors test their approach across several benchmarks under transductive and inductive setups, where it achieves competitive results.

Critical strengths
  1. The paper targets a practical bottleneck in LLM-based KG reasoning, bridging parametric knowledge and structural graph evidence. Generating structural rules to guide reasoning is an interesting approach, particularly for inductive settings involving unseen entities and relations.
  2. The experimental scope is impressive, covering 36 datasets and 17 baselines. The accompanying ablations, backbone comparisons, and efficiency breakdowns provide good empirical support for the framework.
Critical weaknesses
  1. The paper does not clearly explain how reasoning faithfulness is guaranteed in practice or how it is quantitatively evaluated.
  2. The non-triviality of Proposition 1 is insufficiently justified, as it remains unclear why semantic drift reduction is non-trivial if the structural subspace already contains memory representations by definition.
Questions for authors
  1. How do you guarantee and measure reasoning faithfulness in practice?
  2. In Proposition 1, why is the semantic drift reduction non-trivial if the structural subspace already contains the memory representations by definition?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers, ACs, SACs, and PCs

We sincerely thank all reviewers as well as ACs, SACs, and PCs for their careful handling of our submission. We also appreciate the positive feedback from the reviewers on research motivation and practical relevance (LFt6), model design (LFt6, 96GR, and Y6ye), reasoning reliability and interpretability (96GR), and extensive experiments and generalization evaluation (LFt6, 96GR, and Y6ye). Based on the constructive suggestions provided by the reviewers, we have prepared point-by-point responses below. All revisions are highlighted in blue in the updated PDF. Here, we highlight our major responses and hope they can properly address the reviewers' concerns.

  1. Clarification of faithfulness evaluation and the role of RCMP (Reviewers LFt6, 96GR, and Y6ye):
    • Definition and evaluation of faithfulness: We clarify that our objective is to ground reasoning in latent KG structural semantics rather than enforce strict rule execution. We introduce PCA Confidence and Rule-Path Alignment@K, together with Hit@10, to evaluate rule support in observed KGs and its association with entity prediction performance (Section 5.6 and Appendix J.5).
    • Effectiveness of RCMP: We explain how generated rules serve as structural-semantic anchors for relation- and entity-level message passing. The original -RCMP and newly added SRM Init ablations demonstrate the contribution of generated rule semantics beyond additional structural representations. The two-stage WL analysis in Appendix E further supports the enhanced structural distinguishability and expressive power of RCMP. Detailed analysis is highlighted in blue in Appendix J.5.
  2. Methodological Contributions and Comparisons with Existing Methods (Reviewer Y6ye):
    • Overall difference in research motivations: Unlike KGFM and LLM-based KG reasoners, which focus on transferable structural representation learning and KG-LLM knowledge integration, respectively, SIRLM aims to improve the structural-semantic faithfulness of LLM-based KG reasoning, addressing the reasoning evidence perception drift of LLMs over KGs.
    • Methodological Comparison with KGFM: While ULTRA and MOTIF focus on learning transferable structural representations, SIRLM leverages LLMs to generate structural rules and introduces RCMP to incorporate rule semantics into structural message passing, thereby enhancing structural reasoning. The results in Table 1 and the theoretical analysis in Appendix E support the effectiveness of this design.
    • Methodological Comparison with LLM-based KG reasoner: We clarify the methodological novelty in knowledge representation and reasoning mechanisms: ChatRule uses textualized rule generation, and KRLM and FtG focus on LLM-KG knowledge fusion for implicit entity prediction. In contrast, SIRLM generates structural rules through SIRG and incorporates them into neuro-symbolic reasoning via RCMP. We further conduct matched-backbone comparisons with KRLM and FtG, analyze token overhead and computational costs, and introduce Random SRM and SRM Init ablations to validate the contributions of relation memory in SIRG and generated rule semantics.
  3. Theoretical and Implementation Clarifications (Reviewers LFt6 and 96GR):
    • Training objective (Reviewer 96GR): We correct the error in Eq. (14) and reformulate Eq. (15). The modified content has been highlighted in blue at Eqs. (14) and (15).
    • Data construction (Reviewer 96GR): We clarify target-edge filtering during rule mining and training to prevent leakage (Appendix H). An identity-rule ablation further shows that removing identity-rule supervision brings no consistent performance gains, supporting its retention as a fallback (Appendix J.4).
    • Theoretical formulation and assumptions (Reviewers LFt6 and 96GR): We corrected the definition of the structural subspace in Proposition 1 by excluding memory value vectors and revised its proof accordingly, establishing the non-triviality of semantic drift reduction. We also refined Assumption B2 by explicitly introducing a unique-maximum (tie-free) condition and revised the corresponding scope and proof of Proposition 2(I) to ensure theoretical rigor. The above modifications are highlighted in blue in Appendix D and E.2, respectively.
  4. Experimental Fairness and Reliability (Reviewer 96GR): We recalculate relative gains for each training mode of SIRLM against the strongest baseline. We also conduct three-seed experiments on six representative datasets and report mean ± std for MRR and Hit@10, supporting the reliability of our results.

We appreciate your time and careful consideration of our submission. We hope this summary assists you in navigating our rebuttal situation. The revised PDF containing all these improvements has been uploaded.

Best regards,

on behalf of the authors of Paper 14701

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.