Ensuring the safety of LLM-generated content is essential for real-world deployment.Most existing guardrail models formulate moderation as a fixed binary classification task, implicitly assuming a fixed definition of harmfulness.In practice, enforcement strictnesshow conservatively harmfulness is defined and enforced-varies across platforms and evolves over time, making binary moderators brittle under shifting requirements.We first introduce FlexBench, a strictness-adaptive LLM moderation benchmark that enables controlled evaluation under multiple strictness regimes.Experiments on FlexBench reveal substantial cross-strictness inconsistency in existing moderators: models that perform well under one regime can degrade substantially under others, limiting their practical usability.To address this, we propose FlexGuard, an LLM-based moderator that outputs a calibrated continuous risk score reflecting risk severity and supports strictness-specific decisions via thresholding.We train FlexGuard via risk-alignment optimization to improve score-severity consistency and provide practical threshold selection strategies to adapt to target strictness at deployment.Experiments on FlexBench and public benchmarks demonstrate that FlexGuard achieves higher moderation accuracy and substantially improved robustness under varying strictness.

Organic Solar Cells (OSCs) are a promising technology for sustainable energy production. However, the identification of molecules with desired OSC properties typically involves laborious experimental research. To accelerate progress in the field, it is crucial to develop machine learning models capable of accurately predicting the properties of OSC molecules. While graph representation learning has demonstrated success in molecular property prediction, it remains underexplored for OSC-specific tasks. Existing methods fail to capture the unique structural features of OSC molecules, particularly the intricate ring systems that critically influence OSC properties, leading to suboptimal performance. To fill the gap, we present RingFormer, a novel graph transformer framework specially designed to capture both atom and ring level structural patterns in OSC molecules. RingFormer constructs a hierarchical graph that integrates atomic and ring structures and employs a combination of local message passing and global attention mechanisms to generate expressive graph representations for accurate OSC property prediction. We evaluate RingFormer's effectiveness on five curated OSC molecule datasets through extensive experiments. The results demonstrate that RingFormer consistently outperforms existing methods, achieving a 22.77% relative improvement over the nearest competitor on the CEPDB dataset.

Graph Neural Networks (GNNs) are pivotal in graph-based learning, particularly excelling in node classification. However, their scalability is hindered by the need for multi-hop data during inference, limiting their application in latency-sensitive scenarios. Recent efforts to distill GNNs into multi-layer perceptrons (MLPs) for faster inference often underutilize the layer-level insights of GNNs. In this paper, we present TINED, a novel approach that distills GNNs to MLPs on a layer-by-layer basis using Teacher Injection and Dirichlet Energy Distillation techniques. We focus on two key operations in GNN layers: feature transformation (FT) and graph propagation (GP). We recognize that FT is computationally equivalent to a fully-connected (FC) layer in MLPs. Thus, we propose directly transferring teacher parameters from an FT in a GNN to an FC layer in the student MLP, enhanced by fine-tuning. In TINED, the FC layers in an MLP replicate the sequence of FTs and GPs in the GNN. We also establish a theoretical bound for GP approximation. Furthermore, we note that FT and GP operations in GNN layers often exhibit opposing smoothing effects: GP is aggressive, while FT is conservative. Using Dirichlet energy, we develop a DE ratio to measure these effects and propose Dirichlet Energy Distillation to convey these characteristics from GNN layers to MLP layers. Extensive experiments show that TINED outperforms GNNs and leading distillation methods across various settings and seven datasets. Source code are available at https://github.com/scottjiao/TINED_ICML25/.

Equivariant Graph Neural Networks (EGNNs) have demonstrated significant success in modeling microscale systems, including those in chemistry, biology and materials science. However, EGNNs face substantial computational challenges due to the high cost of constructing edge features via spherical tensor products, making them almost impractical for large-scale systems. To address this limitation, we introduce E2Former, an equivariant and efficient transformer architecture that incorporates a Wigner $6j$ convolution (Wigner $6j$ Conv). By shifting the computational burden from edges to nodes, Wigner $6j$ Conv reduces the complexity from $O(| \mathcal{E}|)$ to $O(| \mathcal{V}|)$ while preserving both the model's expressive power and rotational equivariance. We show that this approach achieves a 7x–30x speedup compared to conventional $\mathrm{SO}(3)$ convolutions. Furthermore, our empirical results demonstrate that the derived E2Former mitigates the computational challenges of existing approaches without compromising the ability to capture detailed geometric information. This development could suggest a promising direction for scalable molecular modeling.

BACKGROUND: Machine learning and deep learning are powerful tools for analyzing electronic health records (EHRs) in healthcare research. Although family health history has been recognized as a major predictor for a wide spectrum of diseases, research has so far adopted a limited view of family relations, essentially treating patients as independent samples in the analysis. METHODS: To address this gap, we present ALIGATEHR, which models inferred family relations in a graph attention network augmented with an attention-based medical ontology representation, thus accounting for the complex influence of genetics, shared environmental exposures, and disease dependencies. RESULTS: Taking disease risk prediction as a use case, we demonstrate that explicitly modeling family relations significantly improves predictions across the disease spectrum. We then show how ALIGATEHR's attention mechanism, which links patients' disease risk to their relatives' clinical profiles, successfully captures genetic aspects of diseases using longitudinal EHR diagnosis data. Finally, we use ALIGATEHR to successfully distinguish the 2 main inflammatory bowel disease subtypes with highly shared risk factors and symptoms (Crohn's disease and ulcerative colitis). CONCLUSION: Overall, our results highlight that family relations should not be overlooked in EHR research and illustrate ALIGATEHR's great potential for enhancing patient representation learning for predictive and interpretable modeling of EHRs.

Tables are a prevalent format for structured data, yet their metadata, such as semantic types and column relationships, is often incomplete or ambiguous. Column annotation tasks, including Column Type Annotation (CTA) and Column Property Annotation (CPA), address this by leveraging table context, which are critical for data management. Existing methods typically serialize all columns in a table into pretrained language models to incorporate context, but this coarse-grained approach often degrades performance in wide tables with many irrelevant or misleading columns. To address this, we propose a novel retrieve-and-verify context selection framework for accurate column annotation, introducing two methods: REVEAL and REVEAL+. In REVEAL, we design an efficient unsupervised retrieval technique to select compact, informative column contexts by balancing semantic relevance and diversity, and develop context-aware encoding techniques with role embeddings and target-context pair training to effectively differentiate target and context columns. To further improve performance, in REVEAL+, we design a verification model that refines the selected context by directly estimating its quality for specific annotation tasks. To achieve this, we formulate a novel column context verification problem as a classification task and then develop the verification model. Moreover, in REVEAL+, we develop a top-down verification inference technique to ensure efficiency by reducing the search space for high-quality context subsets from exponential to quadratic. Extensive experiments on six benchmark datasets demonstrate that our methods consistently outperform state-of-the-art baselines.

Large Language Models (LLMs) have demonstrated exceptional performance in biochemical tasks, especially the molecule caption translation task, which aims to bridge the gap between molecules and natural language texts. However, previous methods in adapting LLMs to the molecule-caption translation task required extra domain-specific pre-training stages, suffered weak alignment between molecular and textual spaces, or imposed stringent demands on the scale of LLMs. To resolve the challenges, we propose In-Context Molecule Adaptation (ICMA), as a new paradigm allowing LLMs to learn the molecule-text alignment from context examples via In-Context Molecule Tuning. Specifically, ICMA incorporates the following three stages: Hybrid Context Retrieval, Post-retrieval Re-ranking, and In-context Molecule Tuning. Initially, Hybrid Context Retrieval utilizes BM25 Caption Retrieval and Molecule Graph Retrieval to retrieve similar informative context examples. Additionally, Post-retrieval Re-ranking is composed of Sequence Reversal and Random Walk selection to further improve the quality of retrieval results. Finally, In-Context Molecule Tuning unlocks the in-context learning and reasoning capability of LLMs with the retrieved examples and adapts the parameters of LLMs for better alignment between molecules and texts. Experimental results demonstrate that ICMA can empower LLMs to achieve state-of-the-art or comparable performance without extra training corpora and intricate structures, showing that LLMs are inherently in-context molecule learners.

Cryptocurrencies are rapidly expanding and becoming vital in digital financial markets. However, the rise in cryptocurrency-related illicit activities has led to significant losses for users. To protect the security of these platforms, it is critical to identify illicit accounts effectively. Current detection methods mainly depend on feature engineering or are inadequate to leverage the complex information within cryptocurrency transaction networks, resulting in suboptimal performance. In this paper, we present DIAM, an effective method for detecting illicit accounts in cryptocurrency transaction networks modeled by directed multi-graphs with attributed edges. DIAM first features an Edge2Seq module that captures intrinsic transaction patterns from parallel edges by considering edge attributes and their directed sequences, to generate effective node representations. Then in DIAM, we design a multigraph Discrepancy (MGD) module with a tailored message passing mechanism to capture the discrepant features between normal and illicit nodes over the multigraph topology, assisted by an attention mechanism. DIAM integrates these techniques for end-to-end training to detect illicit accounts from legitimate ones. Extensive experiments, comparing against 15 existing solutions on 4 large cryptocurrency datasets of Bitcoin and Ethereum, demonstrate that DIAM consistently outperforms others in accurately identifying illicit accounts. For example, on a Bitcoin dataset with 20 million nodes and 203 million edges, DIAM attains an F1 score of 96.55%, markedly surpassing the runner-up's score of 83.92%. The code is available at https://github.com/TommyDzh/DIAM.

Graph-level representation learning is important in a wide range of applications. Existing graph-level models are generally built on i.i.d. assumption for both training and testing graphs. However, in an open world, models can encounter out-of-distribution (OOD) testing graphs that are from different distributions unknown during training. A trustworthy model should be able to detect OOD graphs to avoid unreliable predictions, while producing accurate in-distribution (ID) predictions. To achieve this, we present SGOOD, a novel graph-level OOD detection framework. We find that substructure differences commonly exist between ID and OOD graphs, and design SGOOD with a series of techniques to encode task-agnostic substructures for effective OOD detection. Specifically, we build a super graph of substructures for every graph, and develop a two-level graph encoding pipeline that works on both original graphs and super graphs to obtain substructure-enhanced graph representations. We then devise substructure-preserving graph augmentation techniques to further capture more substructure semantics of ID graphs. Extensive experiments against 11 competitors on numerous graph datasets demonstrate the superiority of SGOOD, often surpassing existing methods by a significant margin. The code is available at https://github.com/TommyDzh/SGOOD.

In modern online services, it is of growing importance to process web-scale graph data and high-dimensional sparse data together into embeddings for downstream tasks, such as recommendation, advertisement, prediction, and classification. There exist learning methods and systems for either high-dimensional sparse data or graphs, but not both. There is an urgent need in industry to have a system to efficiently process both types of data for higher business value, which however, is challenging. The data in Tencent contains billions of samples with sparse features in very high dimensions, and graphs are also with billions of nodes and edges. Moreover, learning models often perform expensive operations with high computational costs. It is difficult to store, manage, and retrieve massive sparse data and graph data together, since they exhibit different characteristics. We present EmbedX, an industrial distributed learning framework from Tencent, which is versatile and efficient to support embedding on both graphs and high-dimensional sparse data. EmbedX consists of distributed server layers for graph and sparse data management, and optimized parameter and graph operators, to efficiently support 4 categories of methods, including deep learning models on high-dimensional sparse data, network embedding methods, graph neural networks, and in-house developed joint learning models on both types of data. Extensive experiments on massive Tencent data and public data demonstrate the superiority of EmbedX. For instance, on a Tencent dataset with 1.3 billion nodes, 35 billion edges, and 2.8 billion samples with sparse features in 1.6 billion dimension, EmbedX performs an order of magnitude faster for training and our joint models achieve superior effectiveness. EmbedX is deployed in Tencent. A/B test on real use cases further validates the power of EmbedX. EmbedX is implemented in C++ and open-sourced at https://github.com/Tencent/embedx.

Root Cause Analysis (RCA) is a critical technology for yield improvement in integrated circuit manufacture. Traditional RCA prefers unsupervised algorithms such as Expectation Maximization based on Bayesian models. However, these methods are severely limited by the weak predictive capability of statistical models and can’t effectively transfer the yield learning experience from old designs and processes to the new ones. Motivated by recent advancements of deep learning, in this paper we propose a Neural-Network-based adaptive framework for RCA in yield improvement. The proposed framework consists of an inference module and a self-adaptive module. The former receives volume diagnosis reports and predicts the root cause distributions. The latter is able to adapt the inference module to new designs and processes based on a few of targeted samples without any manual adjustment. Experimental results show that a relatively large improvement on accuracy is achieved by the proposed framework on simulated diagnosis data. Furthermore, the transferring capability of the self-adaptive module is also validated by the results.