📝 Abstract:
Recent research has led to remarkable progress in the capabilities of language models. Current frontier models can tackle very challenging math and coding tasks such as AIME or IOI problems at the cost of generating long chain-of-thought traces (potentially 1000s of tokens). The current need for large models generating many tokens motivates the efficiency question: To what extent can we push the boundaries with smaller models generating fewer tokens? How to attain the pareto-frontier of accuracy-efficiency tradeoff? This talk presents our recent works tackling these. Firstly, I will discuss how suitable use of curriculum learning can help language models solve difficult reasoning problems that are otherwise unsolvable with conventional SFT+RL training. Secondly, I will discuss how language models can benefit from "thinking in a continuous space" by generating superpositions of words rather than limiting themselves to discrete sampling and how it benefits CoT. Paper 1: https://openreview.net/forum?id=NUDaln2vCe Paper 2: https://openreview.net/forum?id=sTPKDKn5ig
👥 Biography:
Samet Oymak is an associate professor of Electrical Engineering and Computer Science at the University of Michigan. His research focuses on optimization theory, statistical learning, decision making, and trustworthy and efficient AI/ML methods. Prior to UMich, he was with the ECE department at the University of California, Riverside. He has also spent time in the finance and tech industry as a researcher and did a postdoc at UC Berkeley as a Simons Fellow. He obtained his PhD degree from Caltech in 2015 for which he received a Charles Wilts Prize for the best departmental thesis. He is a recipient of an NSF CAREER award; faculty research awards from Google, Adobe, Amazon; and an Outstanding Achievement Award from the U of M.