ACL2025 Graph Learning AI paper notes paper summaries Question Answering Reasoning LLM RAG GNNs Multimodal/VLM

🕸️ Graph Learning¶

💬 ACL2025 · 24 paper notes

📌 Same area in other venues: 📷 CVPR2026 (8) · 🔬 ICLR2026 (118) · 💬 ACL2026 (24) · 🧪 ICML2026 (35) · 🤖 AAAI2026 (37) · 🧠 NeurIPS2025 (54)

🔥 Top topics: Question Answering ×6 · Reasoning ×4 · LLM ×4 · RAG ×3 · GNNs ×2

A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning: This paper proposes the Deep Generative Adaptive Replay (DGAR) method, which utilizes a pre-trained diffusion model to generate historical entity distribution representations, mitigates distribution conflicts by enhancing shared features between the historical and current distributions, and designs a layer-wise adaptive replay mechanism to integrate historical and current knowledge, significantly alleviating the catastrophic forgetting problem in continual learning scenarios for temporal knowledge graph reasoning.
A Mutual Information Perspective on Knowledge Graph Embedding: This paper proposes a Knowledge Graph Embedding (KGE) framework based on mutual information maximization. It enhances the semantic representation capability of entities and relations by maximizing the mutual information between different components of triples, achieving consistent performance improvements under complex relational patterns (e.g., 1-N, N-1).
Agent Steerable Search for Knowledge Graph Question Answering: This paper proposes an agent-based steerable knowledge graph search framework, enabling LLM agents to dynamically adjust graph search strategies (such as search depth, direction, and pruning rules) based on question types and reasoning requirements, achieving fine-grained control over the knowledge graph question answering process.
Beyond Completion: A Foundation Model for General Knowledge Graph Reasoning: This paper proposes MERRY, a foundation model for knowledge graphs (KGs) that unifiedly handles both in-KG (zero-shot KGC) and out-of-KG (KGQA) reasoning tasks. By fusing textual and structural information via multi-view conditional message passing (CMP), MERRY outperforms existing methods across 28 datasets.
Can Graph Neural Networks Learn Language with Extremely Weak Text Supervision?: This paper proposes Morpher, a multimodal prompt learning paradigm. Under extremely weak text supervision (only a few tokens of label names), Morpher aligns a pre-trained GNN into the semantic space of an LLM by simultaneously learning graph prompts and text prompts, enabling cross-task and cross-domain graph classification transfer, as well as the first CLIP-style zero-shot GNN classification prototype.
Croppable Knowledge Graph Embedding: Proposes the MED framework to train "croppable" knowledge graph embeddings—optimizing 64 sub-models of different dimensions (sharing embedding prefixes) simultaneously in a single training run. Through mutual learning, evolutionary improvement, and dynamic loss weights, sub-models of each dimension can be directly cropped and used, outperforming independent training and distillation methods while being 10 times faster to train.
Cross-Document Contextual Coreference Resolution in Knowledge Graphs: Proposes a knowledge graph-based cross-document coreference resolution method. By associating textual entity mentions with knowledge graph nodes through a dynamic linking mechanism, it combines contextual embeddings and graph message passing reasoning to improve the precision and recall of cross-document entity recognition, outperforming traditional methods on multiple benchmark datasets.
Disentangled Multi-span Evolutionary Network against Temporal Knowledge Graph Reasoning: The authors propose DiMNet, which separates the active/stable features of node semantics through a multi-span evolution strategy and a cross-time disentanglement mechanism. This significantly improves extrapolation reasoning performance on Temporal Knowledge Graphs (TKGs), achieving SOTA results across four benchmark datasets.
Extending Complex Logical Queries on Uncertain Knowledge Graphs: This paper proposes a formal framework of "soft queries" to extend complex logical queries to uncertain knowledge graphs containing confidence values, and designs the SRC method combining forward reasoning and backward calibration to answer soft queries efficiently, with theoretical proofs that errors do not catastrophically cascade.
Fast-and-Frugal Text-Graph Transformers are Effective Link Predictors: This paper proposes the Fast-and-Frugal Text-Graph (FnF-TG) Transformer, which uniformly encodes textual descriptions and graph structure (ego-graph) via the Transformer's self-attention mechanism. It outperforms SOTA models using large BERT and MPNNs on inductive link prediction tasks using only a small BERT, while extending to a fully inductive setting (where relations can also be inductive) for the first time.
FiDeLiS: Faithful Reasoning in Large Language Model for Knowledge Graph Question Answering: Proposes the FiDeLiS framework, which narrows down the search space via a Path-RAG-preselected candidate set, and progressively constructs and validates reasoning paths using Deductive-Verification Beam Search (DVBS). This improves LLM accuracy and interpretability in knowledge graph question answering without requiring training.
GraphCheck: Breaking Long-Term Text Barriers with Extracted Knowledge Graph-Powered Fact-Checking: GraphCheck proposes a graph-enhanced fact-checking framework that leverages LLMs to extract knowledge graph triples from documents and claims. These triples are encoded by GNNs as graph structures and injected as soft prompts into a frozen LLM validator. This achieves fine-grained fact-checking in a single inference call, yielding an average improvement of \(7.1\%\) across 7 benchmarks and showing strong generalization capability in the medical domain.
GraphNarrator: Generating Textual Explanations for Graph Neural Networks: GraphNarrator is the first method to generate natural language explanations for Graph Neural Networks (GNNs). By utilizing saliency graph verbalization for pseudo-label generation, information-theoretic metric-driven expert iteration for self-improvement, and knowledge distillation for training an end-to-end explainer, it achieves faithful, concise, and human-friendly explanations of GNN decisions.
Can Knowledge Graphs Make Large Language Models More Trustworthy? An Empirical Study Over Open-ended Question Answering: This work proposes OKGQA, an open-ended knowledge graph question answering benchmark, and its perturbed variant OKGQA-P. Through a unified graph-guided retrieval-generation framework, it systematically demonstrates that KG augmentation effectively reduces LLM hallucination rates (boosting FActScore by ~20 percentage points), with subgraph retrieval achieving optimal performance across all query types and exhibiting robustness to KG noise.
Knowledge Graph Retrieval-Augmented Generation for LLM-based Recommendation (K-RagRec): The K-RagRec framework is proposed, which provides structured and reliable external knowledge for LLM-based recommender systems by retrieving multi-hop subgraphs from knowledge graphs. Combining a popularity-based selective retrieval strategy and a GNN encoder, it effectively mitigates hallucination and knowledge deficit issues in LLM recommendations.
M3HG: Multimodal, Multi-scale, and Multi-type Node Heterogeneous Graph for Emotion Cause Triplet Extraction in Conversations: This paper proposes the M3HG model, which explicitly models emotional and causal contexts in conversations by constructing a multimodal multi-type node heterogeneous graph. It fuses semantic information at both intra-utterance and inter-utterance scales to achieve end-to-end extraction of emotion-cause triplets in multimodal conversations. Additionally, it constructs MECAD, the first Chinese multi-scenario MECTEC dataset.
mRAKL: Multilingual Retrieval-Augmented Knowledge Graph Construction for Low-Resourced Languages: Reframe multilingual knowledge graph construction (mKGC) as a question answering (QA) task, and propose mRAKL, a RAG-based system that leverages unstructured monolingual data as a retrieval source to overcome the scarcity of structured data in low-resource languages. The method significantly outperforms existing approaches on two low-resource languages, Tigrinya and Amharic.
Multimodal Transformers are Hierarchical Modal-wise Heterogeneous Graphs: Proving from a graph theory perspective that Multimodal Transformers (MulTs) are essentially Hierarchical Modal-wise Heterogeneous Graphs (HMHGs), this paper proposes the GsiT model. By employing an Interlaced Mask mechanism, GsiT achieves All-Modal-In-One fusion with only 1/3 of the parameters while significantly outperforming traditional MulTs.
Ontology-Guided Reverse Thinking Makes Large Language Models Stronger on Knowledge Graph Question Answering: Proposes ORT (Ontology-Guided Reverse Thinking), which leverages the ontology structure of Knowledge Graphs to construct label reasoning paths backward from the target to guide knowledge retrieval, significantly enhancing the KGQA capabilities of LLMs.
Can LLMs Evaluate Complex Attribution in QA? Automatic Benchmarking using Knowledge Graphs: Proposes the CAQA benchmark, which leverages knowledge graphs to automatically generate a large-scale QA attribution evaluation dataset (161K samples) containing four attribution categories (Supporting, Partially Supporting, Contradictory, Irrelevant) and four reasoning complexity levels. Systematically evaluates 25 automatic attribution evaluators, revealing that "partially supporting" identification and complex reasoning scenarios are the core bottlenecks of current evaluators.
Predicate-Conditional Conformalized Answer Sets for Knowledge Graph Embeddings: This paper proposes CondKGCP, a predicate-conditional conformal prediction method for uncertainty quantification in knowledge graph embeddings. By merging similar predicates to expand the calibration set and employing dual calibration (score + rank) to reduce prediction set size, it outputs tighter answer sets while guaranteeing predicate-level conditional coverage, outperforming five baselines on multiple KGE benchmarks.
RSCF: Relation-Semantics Consistent Filter for Entity Embedding of Knowledge Graph: Proposed RSCF, a plug-and-play KGE method, which ensures "relation-semantics consistency" (semantically similar relations generate similar entity transformations) through three key designs: shared affine transformation, rooted entity transformation, and normalization. It significantly outperforms SOTA models on both distance-based and tensor factorization models, with consistency preservation rates validated theoretically and experimentally.
SimGRAG: Leveraging Similar Subgraphs for Knowledge Graphs Driven Retrieval-Augmented Generation: This paper proposes SimGRAG, an approach that employs a two-stage "query → pattern graph → subgraph" alignment strategy. By leveraging LLMs to convert queries into graph patterns and utilizing Graph Semantic Distance (GSD) to efficiently retrieve the most semantically similar subgraphs in KGs, SimGRAG achieves plug-and-play KG-driven RAG, outperforming all existing methods on question-answering and fact verification tasks.
The Role of Exploration Modules in Small Language Models for Knowledge Graph Question Answering: This paper systematically diagnoses the root cause of why Small Language Models (SLMs, 0.5B–8B) fail under the Think-on-Graph (ToG) knowledge graph question answering framework, revealing that the exploration phase rather than the reasoning phase is the performance bottleneck. It demonstrates that replacing SLM-based KG traversal with a zero-shot, plug-and-play, lightweight sentence retrieval module (SentenceBERT/GTR, only ~110M parameters) consistently and significantly improves Exact Match (EM) scores on both the CWQ and WebQSP benchmarks.