🕸️ Graph Learning¶

💬 ACL2026 · 24 paper notes

📌 Same area in other venues: 📷 CVPR2026 (8) · 🔬 ICLR2026 (118) · 🧪 ICML2026 (35) · 🤖 AAAI2026 (37) · 🧠 NeurIPS2025 (54) · 📹 ICCV2025 (1)

🔥 Top topics: RAG ×5 · Reasoning ×4 · Agents ×2

AgentGL: Towards Agentic Graph Learning with LLMs via Reinforcement Learning: AgentGL is proposed as the first reinforcement learning-based Agentic Graph Learning (AGL) framework. It enables LLM agents to autonomously navigate Text-Attributed Graphs (TAGs) using graph-native search tools, achieving absolute accuracy improvements of up to 17.5% in node classification and 28.4% in link prediction.
ARK: Answer-Centric Retriever Tuning via KG-augmented Curriculum Learning: The ARK framework is proposed, which filters positive samples through a three-dimensional answer sufficiency score (Forward + Backward + Retriever alignment) and utilizes LLM-constructed Knowledge Graphs (KG) to generate hard negative samples of progressive difficulty for curriculum contrastive learning. It achieves an average F1 improvement of 14.5% across 10 datasets.
Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval: This paper proposes ARK: a training-free Knowledge Graph (KG) retrieval agent that exposes only two minimal tools—"global lexical search" and "single-hop neighbor expansion"—allowing the LLM to autonomously switch between breadth and depth without seed nodes or fixed hop counts. It pushes the average Hit@1 on three STaRK graphs to 59.1%, achieving up to a 31.4% improvement over training-free baselines, and enables label-free strategy distillation into Qwen3-8B.
AutoPKG: An Automated Framework for Dynamic E-commerce Product-Attribute Knowledge Graph Construction: AutoPKG is proposed as a multi-agent LLM framework for automatically constructing a Product-Attribute Knowledge Graph (PKG) from multimodal e-commerce content. Using a Type Induction Agent, Attribute Key Discovery Agent, Attribute Value Extraction Agent, and a centralized KGD decision agent, it enables continuous evolution and normalization of a dynamic ontology. It achieves 0.953 WKE (Type) and 0.724 WKE (Key) on the Lazada dataset, with a 7.89% recommendation GMV gain in online A/B testing.
CoG: Controllable Graph Reasoning via Relational Blueprints and Failure-Aware Refinement over Knowledge Graphs: CoG is a training-free KGQA framework that applies Kahneman's Dual-Process Theory to KG reasoning: System 1 distills SPARQL from the training set offline into a "Relational Blueprint" template library, which serves as a soft structural constraint online to guide the reranking and pruning of candidate relations; System 2 triggers evidence-conditioned reflection and targeted backtracking when search stalls, correcting early erroneous decisions. It achieves SOTA accuracy on three multi-hop KGQA benchmarks (GPT-4 backbone: CWQ 77.8, WebQSP 89.7, GrailQA 86.4) while maintaining lower costs (CWQ requires 13% fewer tokens and 12% fewer calls than PoG).
Collaboration of Fusion and Independence: Hypercomplex-driven Robust Multi-Modal Knowledge Graph Completion: M-Hyper encodes multi-modal knowledge graph entities into four orthogonal bases of a biquaternion, carrying three independent modalities (Structure/Visual/Textual) and one fused modality respectively. Through the Hamilton product, it simultaneously achieves "modal independence preservation" and "pairwise sufficient interaction," outperforming 18 baselines on DB15K, MKG-W, and MKG-Y datasets with minimal memory usage and training time.
Comparing Human and Large Language Model Interpretation of Implicit Information: This paper proposes the Implicit Information Extraction (IIE) task and an LLM-based three-stage extraction pipeline (Information Extraction → Reasoning Verification → Temporal Analysis). It constructs structured knowledge graphs to represent the implicit meanings of text. Through comparisons with crowdsourced human judgments, it finds that LLMs are more conservative than humans in socially rich contexts, while humans are more conservative in short factual contexts.
ComplianceNLP: Knowledge-Graph-Augmented RAG for Multi-Framework Regulatory Gap Detection: ComplianceNLP is an end-to-end financial regulatory compliance system that constructs a knowledge graph from 12,847 SEC / MiFID II / Basel III regulations to enhance RAG retrieval. Combined with LEGAL-BERT-based multi-task obligation extraction and threshold-scored gap analysis, it outperforms GPT-4o+RAG by 3.5 points on RegObligation / GapBench with an 87.7 F1. It achieves \(2.8\times\) inference acceleration via domain-specific knowledge distillation + Medusa speculative decoding. Over four months of parallel operation, it processed 9,847 updates, reaching a 96.0% recall rate and a 3.1× increase in analyst efficiency.
CRAFTQA: A Code-Driven Adaptive Framework for Complex Structured Data Reasoning: CRAFTQA uses CodeSTEP to generate executable step-by-step Python reasoning code. When predefined operations are insufficient, CRAFT dynamically generates custom functions, significantly enhancing complex structured data QA capabilities across tables, knowledge graphs (KGs), and temporal knowledge graphs (TKGs). The GPT-4o version achieves 76.6% on the complex reasoning Overall metric.
EA-Agent: A Structured Multi-Step Reasoning Agent for Entity Alignment: This paper proposes EA-Agent, which decomposes Entity Alignment (EA) into a structured multi-step reasoning process. By planning and executing a tool pool (triplet selector + alignment tool + reflector), it achieves interpretable alignment decisions. Combined with reward-guided offline policy optimization to continuously improve planning capabilities, it achieves a Hits@1 improvement of up to 3.17% on DBP15K while mitigating efficiency issues caused by redundant triplets.
Evaluating LLMs on Large-Scale Graph Property Estimation via Random Walks: Existing LLM graph reasoning benchmarks are limited to small graphs (20–50 nodes) and assume full visibility. This paper compresses real-world graphs (up to 2.39M nodes) into prompts using "random walk statistics." It proposes EstGraph to evaluate LLMs on four estimation tasks: node/edge count, community count, graph structure, and influential nodes. Findings show that LLMs achieve \(< 20\%\) relative error on medium-scale graphs and effectively identify graph structures.
From Nodes to Narratives: Explaining Graph Neural Networks with LLMs and Graph Context: Ours proposes Gspell, a lightweight post-hoc explanation framework that projects GNN node embeddings into the LLM embedding space and constructs hybrid prompts (soft prompts + text). This enables LLMs to directly reason over GNN internal representations to generate natural language explanations and explanatory subgraphs, achieving a favorable balance between faithfulness and interpretability on Text-Attributed Graphs (TAGs).
Graph-Based Alternatives to LLMs for Human Simulation: This paper proposes GEMS (Graph-basEd Models for Human Simulation), which models closed-form human behavior simulation tasks as link prediction problems on heterogeneous graphs. It matches or exceeds strong LLM baselines across three datasets and three evaluation settings while reducing the number of parameters by three orders of magnitude.
GS-Quant: Granular Semantic and Generative Structural Quantization for Knowledge Graph Completion: GS-Quant quantizes Knowledge Graph (KG) entities into "coarse-to-fine" discrete code sequences. By constraining RQ-VAE with a hierarchical clustering tree, shallow codes encode global categories (e.g., "Person") while deep codes encode fine-grained attributes (e.g., "Artist"). A GPT-style decoder reconstructs both the entity and its ancestors to enforce causal dependencies among codes. These codes are subsequently added to the LLM vocabulary for LoRA fine-tuning, achieving a Hits@1 improvement of 2.2-2.4 points over the SOTA SSQR on WN18RR and FB15k-237.
IndustryAssetEQA: A Neurosymbolic Operational Intelligence System for Embodied Question Answering in Industrial Asset Maintenance: This paper remodels industrial asset maintenance QA as an "embodied decision-making" task. It proposes IndustryAssetEQA, a neurosymbolic system composed of episodic telemetry, FMEA knowledge graphs, parameterized counterfactual risk simulators, provenance verification, and safety gates. On four industrial datasets, it improves structural validity, counterfactual direction accuracy, and explanation entailment by up to 0.51 / 0.47 / 0.64 respectively, while reducing expert-judged severe over-assertions from 28% to 2%.
LegalGraphRAG: Multi-Agent Graph Retrieval-Augmented Generation for Reliable Legal Reasoning: LegalGraphRAG constructs a hierarchical legal graph using fact, ontology, and rule graphs, and utilizes a Researcher-Auditor-Adjudicator multi-agent workflow for retrieval, verification, and adjudication, improving accuracy and evidence traceability in legal judgment generation.
LLMs Underperform Graph-Based Parsers on Supervised Relation Extraction for Complex Graphs: By comparing four LLMs (7B-70B) with a lightweight graph parser (124M parameters) across six relation extraction datasets, this study finds that when the average number of edges in a document's relation graph exceeds approximately 18, the graph parser consistently and significantly outperforms LLMs. On the most complex ERFGC dataset, the F1 gap reaches 13.2 points, revealing fundamental limitations of LLMs in extracting complex linguistic graph structures.
LogosKG: Hardware-Optimized Scalable and Interpretable Knowledge Graph Retrieval: This paper proposes LogosKG, a hardware-aligned knowledge graph retrieval framework. By transforming graph traversal into multiplication operations of ternary sparse matrices (SUB/OBJ/REL), combined with degree-aware graph partitioning, cross-graph routing, and on-demand caching, it achieves scalable and interpretable high-hop retrieval on billion-edge scale KGs using a single device. Downstream KG-LLM interaction experiments reveal the impact of graph topology on LLM diagnostic reasoning.
MegaRAG: Multimodal Knowledge Graph-Based Retrieval Augmented Generation: MegaRAG utilizes MLLMs to perform parallel entity-relation extraction from each page of long documents, merging them into a Multimodal Knowledge Graph (MMKG). It uses a "subgraph-guided" refinement round to complete cross-modal and cross-page relations. Combined with dual-path retrieval and two-stage answer generation, it significantly outperforms GraphRAG/LightRAG/VisRAG, achieving 64.85% accuracy on SlideVQA(2k) (compared to the best baseline of 27.66%).
Overcoming the Impedance Mismatch: A Theoretical Roadmap for Fusing Foundation Models and Knowledge Graphs: This is a purely theoretical position paper: the authors formalize the phenomenon where foundation models (continuous probability spaces) and knowledge graphs (discrete deterministic structures) are difficult to truly integrate as Impedance Mismatch. Using metric embedding theory, they prove mathematical failure upper bounds for three mainstream categories of solutions—ranging from lexical injection to attention-level integration—and propose a theoretical roadmap for the full knowledge lifecycle: "Emergence—Injection—Editing."
STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation: STEM reformulates multi-hop KGQA from step-by-step path searching to a two-stage process: "generating a query schema graph first, then tracing the evidence subgraph by structure." Through semantic-to-structure projection, Triple-GNN global guidance, and structure-matching retrieval, it significantly improves answer accuracy and evidence coverage on WebQSP and CWQ.
TagRAG: Tag-guided Hierarchical Knowledge Graph Retrieval-Augmented Generation: TagRAG replaces expensive entity community partitioning and global graph summarization in GraphRAG with "object tags + domain tag chains." While significantly reducing construction and retrieval costs, it maintains global knowledge integration capabilities and achieves higher win rates than NaiveRAG, GraphRAG, LightRAG, and MiniRAG on four UltraDomain datasets using the small model Qwen3-4B.
What Makes AI Research Replicable? Executable Knowledge Graphs as Scientific Knowledge Representations: This paper proposes Executable Knowledge Graphs (xKG), which organize technical concepts and runnable code snippets from papers into a three-layer Paper-Technique-Code graph structure. Serving as a plug-and-play knowledge base for research replication agents, it achieves a replication score improvement of up to 10.90 percentage points on the PaperBench Code-Dev task across various agents.
Which bird does not have wings: Negative-constrained KGQA with Schema-guided Semantic Matching and Self-directed Refinement: This paper proposes a new task, Negative-constrained Knowledge Graph Question Answering (NEST KGQA), and the NestKGQA dataset. It introduces PyLF, a Python-formatted logic form to clearly express negative constraints, and the CUCKOO framework. By integrating constraint-aware draft generation, schema-guided semantic matching, and self-directed refinement, the framework achieves efficient and precise answering of multi-constraint questions in a few-shot setting.