Understanding LLM Evaluator Behavior: A Structured Multi-Evaluator Framework for Merchant Risk Assessment

Large Language Models (LLMs) are increasingly used as evaluators of reasoning quality, yet their reliability and bias in payments-risk settings remain poorly understood. We introduce a structured multi-evaluator framework for assessing LLM reasoning in Merchant Category Code (MCC)-based merchant risk assessment, combining a five-criterion rubric with Monte-Carlo scoring to evaluate rationale quality and evaluator stability. Five frontier LLMs generate and cross-evaluate MCC risk rationales under attributed and anonymized conditions. To establish a judge-independent reference, we introduce a consensus-deviation metric that eliminates circularity by comparing each judge's score to the mean of all other judges, yielding a theoretically grounded measure of self-evaluation and cross-model deviation. Results reveal substantial heterogeneity: GPT-5.1 and Claude 4.5 Sonnet show negative self-evaluation bias (-0.33, -0.31), while Gemini-2.5 Pro and Grok 4 display positive bias (+0.77, +0.71), with bias attenuating by 25.8 percent under anonymization. Evaluation by 26 payment-industry experts shows LLM judges assign scores averaging +0.46 points above human consensus, and that the negative bias of GPT-5.1 and Claude 4.5 Sonnet reflects closer alignment with human judgment. Ground-truth validation using payment-network data shows four models exhibit statistically significant alignment (Spearman rho = 0.56 to 0.77), confirming that the framework captures genuine quality. Overall, the framework provides a replicable basis for evaluating LLM-as-a-judge systems in payment-risk workflows and highlights the need for bias-aware protocols in operational financial settings.

TransactionGPT

We present TransactionGPT (TGPT), a foundation model for consumer transaction data within one of world's largest payment networks. TGPT is designed to understand and generate transaction trajectories while simultaneously supporting a variety of downstream prediction and classification tasks. We introduce a novel 3D-Transformer architecture specifically tailored for capturing the complex dynamics in payment transaction data. This architecture incorporates design innovations that enhance modality fusion and computational efficiency, while seamlessly enabling joint optimization with downstream objectives. Trained on billion-scale real-world transactions, TGPT significantly improves downstream classification performance against a competitive production model and exhibits advantages over baselines in generating future transactions. We conduct extensive empirical evaluations utilizing a diverse collection of company transaction datasets spanning multiple downstream tasks, thereby enabling a thorough assessment of TGPT's effectiveness and efficiency in comparison to established methodologies. Furthermore, we examine the incorporation of LLM-derived embeddings within TGPT and benchmark its performance against fine-tuned LLMs, demonstrating that TGPT achieves superior predictive accuracy as well as faster training and inference. We anticipate that the architectural innovations and practical guidelines from this work will advance foundation models for transaction-like data and catalyze future research in this emerging field.

SimAug: Enhancing Recommendation with Pretrained Language Models for Dense and Balanced Data Augmentation

Deep Neural Networks (DNNs) are extensively used in collaborative filtering due to their impressive effectiveness. These systems depend on interaction data to learn user and item embeddings that are crucial for recommendations. However, the data often suffers from sparsity and imbalance issues: limited observations of user-item interactions can result in sub-optimal performance, and a predominance of interactions with popular items may introduce recommendation bias. To address these challenges, we employ Pretrained Language Models (PLMs) to enhance the interaction data with textual information, leading to a denser and more balanced dataset. Specifically, we propose a simple yet effective data augmentation method (SimAug) based on the textual similarity from PLMs, which can be seamlessly integrated to any systems as a lightweight, plug-and-play component in the pre-processing stage. Our experiments across nine datasets consistently demonstrate improvements in both utility and fairness when training with the augmented data generated by SimAug. The code is available at https://github.com/YuyingZhao/SimAug.

Enhancing Distribution and Label Consistency for Graph Out-of-Distribution Generalization

To deal with distribution shifts in graph data, various graph out-of-distribution (OOD) generalization techniques have been recently proposed. These methods often employ a two-step strategy that first creates augmented environments and subsequently identifies invariant subgraphs to improve generalizability. Nevertheless, this approach could be suboptimal from the perspective of consistency. First, the process of augmenting environments by altering the graphs while preserving labels may lead to graphs that are not realistic or meaningfully related to the origin distribution, thus lacking distribution consistency. Second, the extracted subgraphs are obtained from directly modifying graphs, and may not necessarily maintain a consistent predictive relationship with their labels, thereby impacting label consistency. In response to these challenges, we introduce an innovative approach that aims to enhance these two types of consistency for graph OOD generalization. We propose a modifier to obtain both augmented and invariant graphs in a unified manner. With the augmented graphs, we enrich the training data without compromising the integrity of label-graph relationships. The label consistency enhancement in our framework further preserves the supervision information in the invariant graph. We conduct extensive experiments on real-world datasets to demonstrate the superiority of our framework over other state-of-the-art baselines.

GAIM: Attacking Graph Neural Networks via Adversarial Influence Maximization

Recent studies show that well-devised perturbations on graph structures or node features can mislead trained Graph Neural Network (GNN) models. However, these methods often overlook practical assumptions, over-rely on heuristics, or separate vital attack components. In response, we present GAIM, an integrated adversarial attack method conducted on a node feature basis while considering the strict black-box setting. Specifically, we define an adversarial influence function to theoretically assess the adversarial impact of node perturbations, thereby reframing the GNN attack problem into the adversarial influence maximization problem. In our approach, we unify the selection of the target node and the construction of feature perturbations into a single optimization problem, ensuring a unique and consistent feature perturbation for each target node. We leverage a surrogate model to transform this problem into a solvable linear programming task, streamlining the optimization process. Moreover, we extend our method to accommodate label-oriented attacks, broadening its applicability. Thorough evaluations on five benchmark datasets across three popular models underscore the effectiveness of our method in both untargeted and label-oriented targeted attacks. Through comprehensive analysis and ablation studies, we demonstrate the practical value and efficacy inherent to our design choices.

SimCE: Simplifying Cross-Entropy Loss for Collaborative Filtering

The learning objective is integral to collaborative filtering systems, where the Bayesian Personalized Ranking (BPR) loss is widely used for learning informative backbones. However, BPR often experiences slow convergence and suboptimal local optima, partially because it only considers one negative item for each positive item, neglecting the potential impacts of other unobserved items. To address this issue, the recently proposed Sampled Softmax Cross-Entropy (SSM) compares one positive sample with multiple negative samples, leading to better performance. Our comprehensive experiments confirm that recommender systems consistently benefit from multiple negative samples during training. Furthermore, we introduce a \underline{Sim}plified Sampled Softmax \underline{C}ross-\underline{E}ntropy Loss (SimCE), which simplifies the SSM using its upper bound. Our validation on 12 benchmark datasets, using both MF and LightGCN backbones, shows that SimCE significantly outperforms both BPR and SSM.

Can One Embedding Fit All? A Multi-Interest Learning Paradigm Towards Improving User Interest Diversity Fairness

Recommender systems (RSs) have gained widespread applications across various domains owing to the superior ability to capture users' interests. However, the complexity and nuanced nature of users' interests, which span a wide range of diversity, pose a significant challenge in delivering fair recommendations. In practice, user preferences vary significantly; some users show a clear preference toward certain item categories, while others have a broad interest in diverse ones. Even though it is expected that all users should receive high-quality recommendations, the effectiveness of RSs in catering to this disparate interest diversity remains under-explored. In this work, we investigate whether users with varied levels of interest diversity are treated fairly. Our empirical experiments reveal an inherent disparity: users with broader interests often receive lower-quality recommendations. To mitigate this, we propose a multi-interest framework that uses multiple (virtual) interest embeddings rather than single ones to represent users. Specifically, the framework consists of stacked multi-interest representation layers, which include an interest embedding generator that derives virtual interests from shared parameters, and a center embedding aggregator that facilitates multi-hop aggregation. Experiments demonstrate the effectiveness of the framework in achieving better trade-off between fairness and utility across various datasets and backbones.

Enhancing Transformers without Self-supervised Learning: A Loss Landscape Perspective in Sequential Recommendation

Transformer and its variants are a powerful class of architectures for sequential recommendation, owing to their ability of capturing a user's dynamic interests from their past interactions. Despite their success, Transformer-based models often require the optimization of a large number of parameters, making them difficult to train from sparse data in sequential recommendation. To address the problem of data sparsity, previous studies have utilized self-supervised learning to enhance Transformers, such as pre-training embeddings from item attributes or contrastive data augmentations. However, these approaches encounter several training issues, including initialization sensitivity, manual data augmentations, and large batch-size memory bottlenecks. In this work, we investigate Transformers from the perspective of loss geometry, aiming to enhance the models' data efficiency and generalization in sequential recommendation. We observe that Transformers (e.g., SASRec) can converge to extremely sharp local minima if not adequately regularized. Inspired by the recent Sharpness-Aware Minimization (SAM), we propose SAMRec, which significantly improves the accuracy and robustness of sequential recommendation. SAMRec performs comparably to state-of-the-art self-supervised Transformers, such as S$^3$Rec and CL4SRec, without the need for pre-training or strong data augmentations.