MMLS 2025

Abstract: It is undeniable that computing research has the power to rapidly reshape the world we live in, and ML is literally proving this point in real time. But it is also true that we often are not aware or cognizant of the positive and negative impacts of our work. In this talk, I argue that we as researchers need to be more accountable for not just our research results, but how they may be used in downstream applications. Recognizing such impacts is arguably a very challenging task itself. Using my own experience in recent adversarial ML projects, I describe the duality of ML’s impact today, both in real harms it has produced via misuse, and in protective benefits it can provide. I share some of the ethical questions we faced when considering the design and deployment of our tools Glaze and Nightshade, and our experiences through this process. Finally, I suggest some takeaways, including possible perspectives on evaluating new research directions, as well as some concrete research questions that offer potential for positive technical and societal impact.

• Manling Li (Northwestern University) – RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning
• Chenxiao Yang (Toyota Technological Institute at Chicago) – PENCIL: Long Thoughts with Short Memory
• Justin Wang (University of Chicago) – ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning
• Zhiqi Gao (University of Wisconsin–Madison) – Theoretical Physics Benchmark (TPBench) - a Dataset and Study of AI Reasoning Capabilities in Theoretical Physics
• Athanasios Glentis (University of Minnesota) – Memory-Efficient LLM Pretraining via Minimalist Optimizer Design
• Shuo Xie (Toyota Technological Institute at Chicago) – Adam Exploits ℓ∞-geometry of Loss Landscape via Coordinate-wise Adaptivity

Sijia Liu – Robust Unlearning for LLMs
As generative AI systems continue to evolve, the ability to selectively remove information from trained models, known as machine unlearning, has become increasingly essential for ensuring regulatory compliance, enforcing ethical constraints, and mitigating the retention of harmful or sensitive content. This talk focuses on a pressing challenge in this space: the robustness of unlearning in large language models (LLMs). We examine how current unlearning methods remain vulnerable to relearning attacks and post-unlearning fine-tuning, where previously removed knowledge can be partially recovered from a small subset of forgotten or auxiliary data. From an optimization perspective, we introduce a novel connection between robust unlearning and sharpness-aware minimization (SAM), showing that promoting flatter loss landscapes through smoothness-based optimization enhances a model’s resistance to relearning. This draws a natural parallel to principles from adversarial robustness. The talk concludes with a discussion of open challenges and future directions for embedding unlearning into the AI lifecycle, ensuring long-term safety, compliance, and trustworthiness across the data, model, and optimization stack.

Zahra Ghodsi – Collaborating with Confidence: Securing Federated Learning Systems
Artificial Intelligence (AI) is increasingly implemented in distributed settings thanks to its ability to process large amounts of data and its power to enable a wide range of applications. Networks of intelligent devices can therefore work collaboratively to facilitate new directions in several domains such as distributed healthcare and transportation. Deploying AI successfully in the distributed or federated setting requires collaboration of a large number of devices which belong to different parties. This collaboration, however, raises security concerns relating to privacy of assets and robustness in the presence of accidental or intentional errors. In this talk, I outline the challenges in developing secure and privacy-preserving federated learning frameworks where the data or even the identity of participants can be sensitive. I highlight the need for designing new holistic solutions where requirements such as privacy and robustness must be simultaneously guaranteed. I conclude by briefly discussing the lessons learned and future research directions.

Han Zhao – Revisiting Scalarization in Multi-Task Learning
Linear scalarization, i.e., combining all loss functions by a weighted sum, has been the default choice in the literature of multi-task learning (MTL) since its inception. In recent years, there has been a surge of interest in developing Specialized Multi-Task Optimizers (SMTOs) that treat MTL as a multi-objective optimization problem. However, it remains open whether there is a fundamental advantage of SMTOs over scalarization. In this talk, I will revisit scalarization from a theoretical perspective. I will be focusing on linear MTL models and studying whether scalarization is capable of fully exploring the Pareto front. Our findings reveal that, in contrast to recent works that claimed empirical advantages of scalarization, when the model is under-parametrized, scalarization is inherently incapable of full exploration, especially for those Pareto optimal solutions that strike the balanced trade-offs between multiple tasks. I will conclude the talk by briefly discussing the extension of our results to general nonlinear neural networks and our recent work on using online Chebyshev scalarization to controllably steer the search of Pareto optimal solutions.

Wei Hu – Abrupt Learning in Transformers
Training Transformers on algorithmic tasks frequently exhibits an intriguing "abrupt learning" phenomenon in their training dynamics: an extended performance plateau followed by a sudden, sharp improvement. In this talk, I will present several empirical observations aiming to uncover universal characteristics and underlying mechanisms behind such dynamics.

Frederic Koehler – On Inductive Bias in Generative Modeling
There has been a lot of work on understanding the inductive bias of learning via gradient descent and related algorithms. For example, many fascinating phenomena have been discovered in supervised settings such as linearized neural networks, matrix factorization, logistic regression, etc. There are, relatively speaking, fewer such examples which have been worked out in the case of generative modeling and density estimation. I will discuss one such example where we were able to rigorously analyze --- for variational autoencoders --- and the role that the data distribution plays in this setting.

Tianhao Wang – Structured Preconditioners in Adaptive Optimization: A Unified Analysis
We present a novel unified analysis for a broad class of adaptive optimization algorithms with structured (e.g., layerwise, diagonal, and kronecker-factored) preconditioners for both online regret minimization and offline convex optimization. Our analysis not only provides matching rate to several important structured preconditioned algorithms including diagonal AdaGrad, full-matrix AdaGrad, and AdaGrad-Norm, but also gives an improved convergence rate for a one-sided variant of Shampoo over that of original Shampoo. Interestingly, more structured preconditioners (e.g., diagonal Adagrad, AdaGrad-Norm which use less space and compute) are often presented as computationally efficient approximations to full-matrix Adagrad, aiming for improved optimization performance through better approximations. Our unified analysis challenges this prevailing view and reveals, perhaps surprisingly, that more structured preconditioners, despite using less space and computation per step, can outperform their less structured counterparts. To demonstrate this, we show that one-sided Shampoo, which is relatively much cheaper than full-matrix AdaGrad could outperform it both theoretically and experimentally.

Abstract: Since the advent of AI, games have served as progress benchmarks, and most real-world settings are imperfect-information games. Meanwhile, imperfect-information variants of chess have existed for over a century, present extreme challenges, and have been the focus of significant AI research. Beyond calculation needed in regular chess, they require reasoning about information gathering, the opponent’s knowledge, signaling, bluffing, etc. The most popular variant, Fog of War (FoW) chess (aka. dark chess) is a recognized challenge problem in AI after superhuman performance was reached in no-limit Texas hold’em poker. We present Obscuro, the first superhuman AI for FoW chess. It introduces advances to search in imperfect-information games, enabling strong, scalable reasoning. Most prior search techniques - such as those used to achieve superhuman play in no-limit Texas hold’em - require the construction of the “common knowledge set” as a first step, making them unusable for games with this much imperfect information. Experiments against the prior state-of-the-art AI and human players - including the world’s best - show that Obscuro is significantly stronger. FoW chess is now the largest (by amount of imperfect information) turn-based game in which superhuman performance has been achieved and the largest game in which imperfect-information search has been successfully applied. This is joint work with my PhD student Brian Hu Zhang.