Showing 14 papers for 2026-07-02
We propose Structure-Aware Optimal Transport for Self-Supervised Continual Graph Learning. Existing methods optimize node-level consistency that ignores global relational structure, causing inter-node correspondences to distort across tasks in a graph sequence. Our structure-aware OT objective aligns relational structures across tasks, improving stability and accuracy in continual graph learning without labels.
We present Multi-Label Node Classification with Label Influence Propagation. The method leverages label co-occurrence and embeddings to propagate label influence over the graph, capturing correlations among labels that standard GNNs may miss. Experiments on MLNC benchmarks demonstrate improved accuracy and label-consistency.
We introduce Graph-Native Reinforcement Learning for traceable scientific hypothesis generation through conceptual recombination (Graph-PRefLexOR family, GRPO). Unlike standard LLMs, our graph-native RL organizes reasoning into explicit phases, yielding more interpretable, domain-grounded hypotheses for materials design. The approach improves traceability of the reasoning process and hypothesis quality.
FedIA proposes Importance-Aware Aggregation for domain-robust Federated Graph Learning. By analyzing the signals exposed to server aggregation after local message passing on domain-separated graph clients, we identify aggregation-level failures and show how to weight updates by domain importance to improve robustness. Empirical results on federated graph benchmarks validate gains over baselines.
We propose Self-Improving Neural Pruning for scalable mixed bundle pricing, using a Graph Neural Network-guided pruning framework to prune the combinatorial bundle space. This reduces the number of bundles to consider while preserving optimality approximations, enabling scalable (non)additive bundle pricing in practice. Experiments show profit gains and faster computation.
Crystalite is a lightweight diffusion Transformer for efficient crystal modeling. It relies on two inductive biases: Subatomic Tokenization, replacing high-dimensional one-hot encodings with compact atom representations, and a Geometry Enhancement Module (GEM) to inject geometric priors. Together they enable efficient training and sampling while preserving crystal structure.
PETIMOT presents a novel SE(3)-equivariant GNN framework to infer protein motions from sparse data. A task-specific loss enforces data symmetries such as scaling and permutation, yielding continuous, physically meaningful representations of protein motions. This enables robust motion inference from limited measurements.
FeLoG offers scalable and efficient distributed graph embedding with a Feedback Loop Mechanism. It couples sampling and training by a feedback loop that aligns sampling decisions with current embedding quality, reducing redundant exploration and improving scalability on large graphs. Experiments demonstrate faster embedding with competitive quality.
This work moves from strings to things for personal knowledge graphs by evaluating LLM-based triple extraction for PKGs used in recommendation systems. We develop a reproducible pipeline to extract RDF-compliant user-preference triples grounded to Wikidata, comparing Qwen- and Gemma-based models. The results inform the viability of building PKGs from conversational data for recommendations.
AGE introduces Adaptive-masking for Graph Embedding to align graph representations with text-based LLMs in Graph Retrieval-Augmented Generation. A Transformer-based masking approach in self-supervised learning adapts graph embeddings for frozen LLMs, improving cross-modal alignment and retrieval performance.
This article surveys or develops optimal any-angle path planning in static and dynamic environments. It focuses on algorithms that guarantee optimal solutions when dynamic obstacles are present, addressing challenges in real-world navigation. It provides theoretical guarantees and practical guidance for implementing optimal any-angle planners.
Diffusion-GR2 is a diffusion-based reasoning re-ranker that speeds up generation by decoding many positions in parallel. It addresses accuracy gaps when converting autoregressive re-rankers to diffusion form by structuring reasoning and aligning with ranking goals. Experiments show faster inference with competitive accuracy.
Hey, That's My Model! Chain & Hash introduces LLM fingerprinting. It defines five properties—transparency, efficiency, persistence, robustness, and unforgeability—and presents chain-and-hash cryptographic fingerprinting to prove ownership and resist tampering.
This work develops multimodal and multiscale spatial-temporal semantic search and recommendation using AI foundation models. The framework leverages LLMs and vision-language models to perform similarity search and ranking for event documents with spatial-temporal information, introducing two new strategies to support effective retrieval and recommendation.