The rapid development of Large Language Models (LLMs) has intensified concerns about content traceability and potential misuse. Existing watermarking schemes for sampled text often face trade-offs between maintaining text quality and ensuring robust detection against various attacks. To address these issues, we propose a novel watermarking scheme that improves both detectability and text quality by introducing a cumulative watermark entropy threshold. Our approach is compatible with and generalizes existing sampling functions, enhancing adaptability. Experimental results across multiple LLMs show that our scheme significantly outperforms existing methods, achieving over 80\% improvements on widely-used datasets, e.g., MATH and GSM8K, while maintaining high detection accuracy.

We present a novel framework that bridges the gap between the interpretability of decision trees and the advanced reasoning capabilities of large language models (LLMs) to predict startup success. Our approach leverages chain-of-thought prompting to generate detailed reasoning logs, which are subsequently distilled into structured, human-understandable logical rules. The pipeline integrates multiple enhancements - efficient data ingestion, a two-step refinement process, ensemble candidate sampling, simulated reinforcement learning scoring, and persistent memory - to ensure both stable decision-making and transparent output. Experimental evaluations on curated startup datasets demonstrate that our combined pipeline improves precision by 54% from 0.225 to 0.346 and accuracy by 50% from 0.46 to 0.70 compared to a standalone OpenAI o3 model. Notably, our model achieves over 2x the precision of a random classifier (16%). By combining state-of-the-art AI reasoning with explicit rule-based explanations, our method not only augments traditional decision-making processes but also facilitates expert intervention and continuous policy refinement. This work lays the foundation for the implementation of interpretable LLM-powered decision frameworks in high-stakes investment environments and other domains that require transparent and data-driven insights.

Hate speech detection is a crucial area of research in natural language processing, essential for ensuring online community safety. However, detecting implicit hate speech, where harmful intent is conveyed in subtle or indirect ways, remains a major challenge. Unlike explicit hate speech, implicit expressions often depend on context, cultural subtleties, and hidden biases, making them more challenging to identify consistently. Additionally, the interpretation of such speech is influenced by external knowledge and demographic biases, resulting in varied detection results across different language models. Furthermore, Large Language Models often show heightened sensitivity to toxic language and references to vulnerable groups, which can lead to misclassifications. This over-sensitivity results in false positives (incorrectly identifying harmless statements as hateful) and false negatives (failing to detect genuinely harmful content). Addressing these issues requires methods that not only improve detection precision but also reduce model biases and enhance robustness. To address these challenges, we propose a novel method, which utilizes in-context learning without requiring model fine-tuning. By adaptively retrieving demonstrations that focus on similar groups or those with the highest similarity scores, our approach enhances contextual comprehension. Experimental results show that our method outperforms current state-of-the-art techniques. Implementation details and code are available at TBD.

--Just-in-time defect prediction (JIT DP) leverages machine learning to identify defect-prone code commits, enabling quality assurance (QA) teams to allocate resources more efficiently by focusing on commits that are most likely to contain defects. Although JIT defect prediction techniques have introduced notable improvements in terms of predictive accuracy, they are still susceptible to misclassification errors such as false positives and false negatives. This can lead to wasted resources or undetected defects, a particularly critical concern when QA resources are limited. T o mitigate these challenges and preserve the practical utility of JIT defect prediction tools, it becomes essential to estimate the reliability of the predictions, i.e., computing confidence scores. Such scores can help practitioners determine the trustworthiness of predictions and and thus prioritize them efficiently. A simple approach to computing confidence scores is to extract, alongside each prediction, the corresponding prediction probabilities and use them as indicators of confidence. However, for these probabilities to reliably serve as confidence scores, the predictive model must be well-calibrated. This means that the prediction probabilities must accurately represent the true likelihood of each prediction being correct. Miscalibration, common in modern machine learning models, distorts probability scores such that the model's prediction probabilities do not align with the actual probability of those predictions being correct. Despite its importance, model calibration has been largely overlooked in JIT defect prediction. In this study, we evaluate the calibration of several state-of-the-art JIT defect prediction techniques to determine whether and to what extent they exhibit poor calibration. Furthermore, we assess whether post-calibration methods can improve the calibration of existing JIT defect prediction models. Our experimental analysis reveals that all evaluated JIT DP models exhibit some level of miscalibration, with Expected Calibration Error (ECE) ranging from 2% to 35%. Furthermore, post-calibration methods do not consistently improve the calibration of these JIT DP models. In recent years, just-in-time defect prediction (JIT DP) has emerged as a valuable machine learning (ML)-based technique, designed to predict whether a code commit is defect-prone or clean. By identifying code commits that are more likely to contain defects, JIT defect prediction helps quality assurance (QA) practitioners decide whether to perform targeted inspections and code reviews, as well as where and how to allocate testing efforts and resources [3], [4]. By supporting the prioritization of the code commits for further investigation and testing, JIT defect prediction models enable the timely identification of defects in the codebase. JIT defect prediction thus provides a means to optimize QA workflows.

Transfer learning from pre-trained encoders has become essential in modern machine learning, enabling efficient model adaptation across diverse tasks. However, this combination of pre-training and downstream adaptation creates an expanded attack surface, exposing models to sophisticated backdoor embeddings at both the encoder and dataset levels--an area often overlooked in prior research. Additionally, the limited computational resources typically available to users of pre-trained encoders constrain the effectiveness of generic backdoor defenses compared to end-to-end training from scratch. In this work, we investigate how to mitigate potential backdoor risks in resource-constrained transfer learning scenarios. Specifically, we conduct an exhaustive analysis of existing defense strategies, revealing that many follow a reactive workflow based on assumptions that do not scale to unknown threats, novel attack types, or different training paradigms. In response, we introduce a proactive mindset focused on identifying clean elements and propose the Trusted Core (T-Core) Bootstrapping framework, which emphasizes the importance of pinpointing trustworthy data and neurons to enhance model security. Our empirical evaluations demonstrate the effectiveness and superiority of T-Core, specifically assessing 5 encoder poisoning attacks, 7 dataset poisoning attacks, and 14 baseline defenses across five benchmark datasets, addressing four scenarios of 3 potential backdoor threats.

We present a refined version of the Anomaly Awareness framework for enhancing unsupervised anomaly detection. Our approach introduces minimal supervision into Variational Autoencoders (VAEs) through a two-stage training strategy: the model is first trained in an unsupervised manner on background data, and then fine-tuned using a small sample of labeled anomalies to encourage larger reconstruction errors for anomalous samples. We validate the method across diverse domains, including the MNIST dataset with synthetic anomalies, network intrusion data from the CICIDS benchmark, collider physics data from the LHCO2020 dataset, and simulated events from the Standard Model Effective Field Theory (SMEFT). The latter provides a realistic example of subtle kinematic deviations in Higgs boson production. In all cases, the model demonstrates improved sensitivity to unseen anomalies, achieving better separation between normal and anomalous samples. These results indicate that even limited anomaly information, when incorporated through targeted fine-tuning, can substantially improve the generalization and performance of unsupervised models for anomaly detection.

Purpose: Accurately classifying tissue margins during cancer surgeries is crucial for ensuring complete tumor removal. Rapid Evaporative Ionization Mass Spectrometry (REIMS), a tool for real-time intraoperative margin assessment, generates spectra that require machine learning models to support clinical decision-making. However, the scarcity of labeled data in surgical contexts presents a significant challenge. This study is the first to develop a foundation model tailored specifically for REIMS data, addressing this limitation and advancing real-time surgical margin assessment. Methods: We propose FACT, a Foundation model for Assessing Cancer Tissue margins. FACT is an adaptation of a foundation model originally designed for text-audio association, pretrained using our proposed supervised contrastive approach based on triplet loss. An ablation study is performed to compare our proposed model against other models and pretraining methods. Results: Our proposed model significantly improves the classification performance, achieving state-of-the-art performance with an AUROC of $82.4\% \pm 0.8$. The results demonstrate the advantage of our proposed pretraining method and selected backbone over the self-supervised and semi-supervised baselines and alternative models. Conclusion: Our findings demonstrate that foundation models, adapted and pretrained using our novel approach, can effectively classify REIMS data even with limited labeled examples. This highlights the viability of foundation models for enhancing real-time surgical margin assessment, particularly in data-scarce clinical environments.

M. Tanveer * Department of Mathematics Indian Institute of T echnology Indore mtanveer@iiti.ac.in Abstract --Restricted kernel machines (RKMs) represent a versatile and powerful framework within the kernel machine family, leveraging conjugate feature duality to address a wide range of machine learning tasks, including classification, regression, and feature learning. However, their performance can degrade significantly in the presence of noise and outliers, which compromises robustness and predictive accuracy. In this paper, we propose a novel enhancement to the RKM framework by integrating a class-informed weighted function. This weighting mechanism dynamically adjusts the contribution of individual training points based on their proximity to class centers and class-specific characteristics, thereby mitigating the adverse effects of noisy and outlier data. By incorporating weighted conjugate feature duality and leveraging the Schur complement theorem, we introduce the class-informed restricted kernel machine (CI-RKM), a robust extension of the RKM designed to improve generalization and resilience to data imperfections. Experimental evaluations on benchmark datasets demonstrate that the proposed CI-RKM consistently outperforms existing baselines, achieving superior classification accuracy and enhanced robustness against noise and outliers. Our proposed method establishes a significant advancement in the development of kernel-based learning models, addressing a core challenge in the field.

We introduce a scalable framework for regressing multivariate distributions onto multivariate distributions, motivated by the application of inferring cell-cell communication from population-scale single-cell data. The observed data consist of pairs of multivariate distributions for ligands from one cell type and corresponding receptors from another. For each ordered pair $e=(l,r)$ of cell types $(l \neq r)$ and each sample $i = 1, \ldots, n$, we observe a pair of distributions $(F_{ei}, G_{ei})$ of gene expressions for ligands and receptors of cell types $l$ and $r$, respectively. The aim is to set up a regression of receptor distributions $G_{ei}$ given ligand distributions $F_{ei}$. A key challenge is that these distributions reside in distinct spaces of differing dimensions. We formulate the regression of multivariate densities on multivariate densities using a generalized Bayes framework with the sliced Wasserstein distance between fitted and observed distributions. Finally, we use inference under such regressions to define a directed graph for cell-cell communications.

Complex networks are frequently employed to model physical or virtual complex systems. When certain entities exist across multiple systems simultaneously, unveiling their corresponding relationships across the networks becomes crucial. This problem, known as network alignment, holds significant importance. It enhances our understanding of complex system structures and behaviours, facilitates the validation and extension of theoretical physics research about studying complex systems, and fosters diverse practical applications across various fields. However, due to variations in the structure, characteristics, and properties of complex networks across different fields, the study of network alignment is often isolated within each domain, with even the terminologies and concepts lacking uniformity. This review comprehensively summarizes the latest advancements in network alignment research, focusing on analyzing network alignment characteristics and progress in various domains such as social network analysis, bioinformatics, computational linguistics and privacy protection. It provides a detailed analysis of various methods' implementation principles, processes, and performance differences, including structure consistency-based methods, network embedding-based methods, and graph neural network-based (GNN-based) methods. Additionally, the methods for network alignment under different conditions, such as in attributed networks, heterogeneous networks, directed networks, and dynamic networks, are presented. Furthermore, the challenges and the open issues for future studies are also discussed.