Two Faces of Optimization:
From Offline Control to Sequential Decisions

Optimization lies at the core of modern machine learning, underpinning both the design of algorithms and the analysis of their performance. This two-day workshop explores recent advances in the theory of optimization from two complementary perspectives. The first day focuses on stochastic approaches, including sampling algorithms, the dynamics of stochastic gradient methods, and probabilistic frameworks with connections to optimal transport. The second day turns to adaptive optimization, highlighting developments in reinforcement learning, bandits, and control, with an emphasis on methodology and fundamental limits.

Speakers

Ayush Sekhari

Massachusetts Institute of Technology

Aditi Krishnapriyan

University of California, Berkeley

Andre Wibisono

Yale University

Matthew Zhang

University of Toronto

Claire Vernade

University of Tübingen

Aldo Pacchiano

Boston University and Broad Institute of MIT and Harvard

Kevin Jamieson

University of Washington

Nicolo Cesa-Bianchi

Università degli Studi di Milano and Politecnico di Milano

Attend and Present

Held at the ETH Zurich OAT, the workshop will be a small gathering accommodating a limited number of attendees. To help us plan, please submit your attendance request using the following form:

Register to Attend

To present a recent work as a poster, you can submit an abstract or a paper. The list of accepted posters will be updated frequently on the website. Posters should be printed at A1 (landscape) or A0 (portrait) for the event.

Submit a Poster

Schedule

Posters

• Exploiting Curvature in Online Convex Optimization with Delayed Feedback
• Optimism in the Face of Ambiguity Principle for Multi-Armed Bandits
• DISCOVER: Automated Curricula for Sparse-Reward Reinforcement Learning
• Efficient Preference-Based Reinforcement Learning: Randomized Exploration Meets Experimental Design
• From Gradient Clipping to Normalization for Heavy Tailed SGD
• Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning
• Safe-EF: Error Feedback for Nonsmooth Constrained Optimization
• ROC-n-reroll: How Verifier Imperfection Affects Test-Time Scaling

Organizers