Mahdi Soltanolkotabi - Interpreting Generative Models for Better Steering: Generation to Verifiable @IPAMUCLA

Uploaded September 2026 | Updated September 2026, 3 weeks ago

Recorded 03 September 2026. Mahdi Soltanolkotabi of the University of Southern California presents "Interpreting Generative Models for Better Steering: From Visual Generation to Verifiable Reasoning" at IPAM's Foundations of Interpretability Workshop.
Abstract: Despite their impressive capabilities, generative models continue to struggle with basic forms of visual reasoning, including reliably generating a specified number of objects. We study the internal dynamics of diffusion models and how visual concepts and scene structure emerge over the denoising trajectory. These insights lead to an early-time steering method that intervenes when global semantic content is first formed. We then study reinforcement learning with verifiable rewards and show how asymmetrically weighting positive and negative feedback can improve visual reasoning. Our results demonstrate how interpretability can provide a principled foundation for steering visual generation and enhancing visual and verifiable reasoning.
Learn more online at: https://www.ipam.ucla.edu/programs/workshops/foundations-of-interpretability/?tab=overview

Mahdi Soltanolkotabi - Interpreting Generative Models for Better Steering: Generation to Verifiable

$Hong Qin - Machine Learning of Local Confinement & Transport from Intl Multi-Tokamak Confinement DB Recorded 06 May 2026. Hong Qin of Princeton University presents Machine Learning of Local Confinement and Transport from the International Multi-Tokamak Confinement Profile Database at IPAMs Fusion Device Design and Engineering Workshop. Abstract: With the accelerated push toward commercial magnetic DT fusion, tritium self-sufficiency has emerged as a critical challenge. One of the key quantities governing tritium self-sufficiency is the tritium burn fraction, which depends critically on particle confinement time in magnetically confined plasmas. Theoretically, particle confinement is controlled by transport coefficients that are functions of plasma and device parameters local to each flux surface. Much of the historical effort to characterize these coefficients has relied on empirical scaling laws, typically developed for individual devices. We present a Gaussian process regression study of local confinement and transport coefficients based on the 2008 public release of the International Multi-Tokamak Confinement Profile Database, which includes 344 discharges from TFTR, JET, DIII-D, Tore Supra, T-10, ITER simulations, JT-60U, ASDEX Upgrade, RTP, MAST, Alcator C-Mod, FTU, and TEXTOR. Preliminary results will be discussed, together with their implications for cross-device confinement modeling. Learn more online at: https://www.ipam.ucla.edu/programs/workshops/workshop-iii-fusion-device-design-and-engineering/$

Syunichi Shiraiwa - Status of physics integration in radio-frequency H/CD modeling - IPAM at UCLA

Florian Schaefer - Information geometric regularization for sensitivities of flows with shocks

Nikola Kovachki - Demystifying Data-Driven Probabilistic Medium-Range Weather Forecasting

Brian Spears - The Genesis Mission: AI for US Science and Security Transformation - IPAM at UCLA

Viktória Földvári - Maximal tb Representatives of Double Twist Knots - IPAM at UCLA

Eduardo Sontag - A Mathematical Model of Evolution of Drug-Induced Resistance - IPAM at UCLA

Ben Raphael - Models and Algorithms for Cancer Evolution, Pt. 2/2 - IPAM at UCLA

Miriam Kuzbary - 0-Surgeries on Links - IPAM at UCLA

Tanushree Shah - Contact cosmetic surgery - IPAM at UCLA

Haiwen Guan - LUCIE Lightweight Uncoupled Climate Emulator long-term stability, physical consistency