We are pleased to announce our upcoming webinar in September 2025. Dr. Jason Klusowski from Princeton University will give a talk at 2pm (ET) on September 24 (Wednesday). Please use the link below to register for the KISS webinar. The webinar title and abstract are as follows.
Date/Time: 2pm – 3pm ET (1pm – 2pm CT; 11am – 12pm PT) on September 24
Registration link:
https://yonsei.zoom.us/meeting/register/USTBSjJESduzP7-IlK15TA
If the link does not open when clicked, please copy and paste it into your browser’s address bar.
Registration is required for this meeting. After registering, you will receive a confirmation email containing information about joining the meeting.
Speaker: Jason Klusowski from Princeton University.
Title: Decoding Game: On Minimax Optimality of Heuristic Text Generation Strategies
Abstract: Decoding strategies play a pivotal role in text generation for modern language models, yet a perplexing gap persists between theory and practice. Surprisingly, strategies that should intuitively be optimal, such as Maximum a Posteriori, often perform poorly in practice. Meanwhile, popular heuristic approaches like Top-k and Nucleus sampling, which employ truncation and renormalization of the conditional next-token probabilities, have achieved great empirical success but lack theoretical justification. This talk introduces the Decoding Game, a theoretical framework that attempts to reconcile this gap by recasting text generation as a two-player zero-sum game. In this game, the Strategist aims to produce text that aligns with the true distribution, while Nature acts as an adversary, distorting the Strategist’s target distribution. Our analysis reveals that the adversarial Nature imposes an implicit regularization on the likelihood, with truncation-renormalization methods emerging as first-order approximations of the optimal strategy.