Machine unlearning is often introduced as efficient data deletion: when specified training records must be removed, update the model to match retraining without them. But language-model unlearning now includes requests such as removing a concept or suppressing a copyrighted work, where neither the relevant training records nor the right retraining counterfactual may be obvious. As unlearning moves beyond record deletion, specifying what to forget becomes part of the technical problem. My recent work studies this shift from three directions: certifying model updates once the forget set is given, selecting data to remove when the target is a distribution, and constructing forget sets from higher-level requests for language models. These results suggest separating three parts of the problem—what should be forgotten, which data should represent that goal, and how the model should be updated—and asking for guarantees that match the intended form of forgetting.

As large language models (LLMs) evolve into multimodal, multi-agent, and conversational systems, privacy risks extend far beyond simple memorization. In this talk, I present a cohesive view of my recent work on understanding and controlling sensitive information leakage in LLMs or agentic systems under adversarial interaction. I begin by examining whether sensitive information can truly be deleted from model weights, showing that even state-of-the-art unlearning and model editing methods leave recoverable traces under both white-box and black-box attacks. I then extend this analysis to multimodal LLMs, introducing UnLOK-VQA, a benchmark for multimodal unlearning, and demonstrating that cross-modal attacks are significantly more effective than text- or image-only attacks. Next, I address the core tension between deletion and utility, presenting UPCORE, a coreset-selection based framework that improves unlearning by selectively pruning forget data to reduce collateral damage. I then move to multi-agent settings, where privacy leakage emerges compositionally across interactions. I show how Theory-of-Mind reasoning and collaborative defenses mitigate these risks. Finally, I introduce the Double-Agent paradigm, where privacy protection becomes a strategic interaction: the defender must model the adversary’s beliefs and deliberately steer them toward a false belief that sensitive information has been extracted. Overall, this talk argues that safeguarding privacy in modern LLM systems requires moving beyond static notions of information deletion toward adversarially robust, interaction-aware, and utility-preserving approaches.

For most of the last decade, machine learning treated reliability as something to be established once: a model was evaluated before deployment, and that evaluation was trusted for its entire service life. Modern AI systems no longer fit this picture. They run continually, accumulate history, and update their own state, so the system in deployment gradually drifts away from the one that was evaluated. This talk asks what it takes to sustain reliability over that lifetime. Drawing on my research, I will discuss how to withstand corruption of an agent’s operating process, how to detect degradation that emerges only along a trajectory, and how to govern what a long-running agent retains and exposes. I will close with an outlook on reliability as something we sustain across a lifetime, rather than establish once and trust.

Evangelos (Vagelis) Papalexakis is a Full Professor and the Ross Family Chair of the CSE Dept. at University of California Riverside. He received his PhD degree at the School of Computer Science at Carnegie Mellon University (CMU). Prior to CMU, he obtained his Diploma and MSc in Electronic & Computer Engineering at the Technical University of Crete, in Greece. Broadly, his research interests span the fields of Data Science, Machine Learning, Artificial Intelligence, and Signal Processing.

His research involves designing interpretable models and scalable algorithms for extracting knowledge from large multi-aspect datasets, with specific emphasis on tensor factorization models, and applying those algorithms to a variety of real-world problems, including AI for Science and Engineering, explainable AI, gravitational wave detection, cybersecurity, transportation and railway safety, and precision agriculture.

He is heavily involved in the data science research community with extensive experience in conference organization, including organizing a workshop at ACM SIGKDD 2019 on "Tensor Methods for Emerging Data Science Problems", being the Deep Learning Day Co-Chair for ACM SIGKDD 2019, the Doctoral Forum Co-Chair for SIAM SDM 2021, the Demos Co-Chair for ACM WSDM 2022, the Program Co-Chair for SIAM SDM 2022, the General Co-Chair for SIAM SDM 2024 and SIAM SDM 2025, the Blue Sky Track Co-Chair for KDD 2026, and is currently serving as the Steering Committee Chair for SIAM SDM.

His work has appeared in top-tier conferences and journals, and has attracted a number of distinctions, including the 2017 SIGKDD Dissertation Award (runner-up), a number of paper awards, the National Science Foundation CAREER award, the 2021 IEEE DSAA Next Generation Data Scientist Award, the 2022 IEEE Signal Processing Society Donald G. Fink Overview Paper Award, and the IEEE ICDM 2022 Tao Li Award and 2025 PAKDD Early Career Research Award, both of which award excellence in early-career researchers in data mining. Finally, Prof. Papalexakis is a 2025 alumnus of the Frontiers of Engineering (FOE) Symposium by the National Academy of Engineering (NAE).

Tensors methods have been traditionally widely used in expressing complex relations in data and extracting interpretable latent insights. At the same time, more recently, tensor methods have been shown to represent model weights, with model compression through low-rank factorization being the most successful such application. In this talk we explore a number of exciting examples where tensor methods, whether applied directly on a model or indirectly in inputs or outputs of a model, can lead to more compact, trustworthy, and robust models.