Talk
Talk by Prof. Subbarao Kambhampati
Title: On the Mythos of LRM "Thinking Tokens"
Abstract:
Ever since DeepSeek R1, the intermediate token output by the Large Reasoning Models (LRMs) have been ascribed several magical properties-- that they provide an interpretable window into the "thinking" of the model, that their length is a reliable indicator of the computational complexity of the task, and that the main contribution of RL-based post-training is to improve model's thinking--as evidenced by the average length of the thinking tokens increasing during training. In this talk, I will discuss several recent efforts from our group that systematically dissect these claims and offer counterintuitive conclusions.
Ever since DeepSeek R1, the intermediate token output by the Large Reasoning Models (LRMs) have been ascribed several magical properties-- that they provide an interpretable window into the "thinking" of the model, that their length is a reliable indicator of the computational complexity of the task, and that the main contribution of RL-based post-training is to improve model's thinking--as evidenced by the average length of the thinking tokens increasing during training. In this talk, I will discuss several recent efforts from our group that systematically dissect these claims and offer counterintuitive conclusions.
Speaker Bio:
Subbarao Kambhampati is a professor of computer science at Arizona State University. Kambhampati studies fundamental problems in planning and decision making, motivated in particular by the challenges of human-aware AI systems. He is a fellow of Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science, and Association for Computing machinery, and a recent recipient of the AAAI Patrick H. Winston Outstanding Educator award. He served as the president of the Association for the Advancement of Artificial Intelligence, a trustee of the International Joint Conference on Artificial Intelligence, the chair of AAAS Section T (Information, Communication and Computation), and a founding board member of Partnership on AI. Kambhampati's research as well as his views on the progress and societal impacts of AI have been featured in multiple national and international media outlets. He can be followed on Twitter click here
Subbarao Kambhampati is a professor of computer science at Arizona State University. Kambhampati studies fundamental problems in planning and decision making, motivated in particular by the challenges of human-aware AI systems. He is a fellow of Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science, and Association for Computing machinery, and a recent recipient of the AAAI Patrick H. Winston Outstanding Educator award. He served as the president of the Association for the Advancement of Artificial Intelligence, a trustee of the International Joint Conference on Artificial Intelligence, the chair of AAAS Section T (Information, Communication and Computation), and a founding board member of Partnership on AI. Kambhampati's research as well as his views on the progress and societal impacts of AI have been featured in multiple national and international media outlets. He can be followed on Twitter click here
- 15 - 15 December,2025
- 11:00 am - 12:00 pm
- 502 Academic Block East (99C)