Keiran Paster

prof_pic.jpg

I am a fifth-year PhD candidate in the Department of Computer Science at the University of Toronto and I am currently on the job market.

My research interests lie in AI reasoning and decision-making and my goal is to create AI systems that effectively scale with compute to process the world’s information and fast-forward scientific progress.

Research highlights from the last year include:

  • Automatic prompt engineering, which shows for the first time that when given access to trial-and-error, LLMs can control themselves through prompting often better than a human can (resulting in the discovery of a prompting trick that gives insight into the psychology of GPT).
  • STEVE-1, an instruction-following agent in Minecraft created using a novel methodology where we finetune a model pretrained on years of Minecraft videos on goal-relabeled data. STEVE-1 acts directly using keyboard and mouse input and follows open-ended text and visual instructions.
  • OpenWebMath, 14.7B tokens of mathematical documents gathered from Common Crawl for use in LLM pretraining and midtraining. OpenWebMath improves mathematical reasoning performance over 20x more effectively per-token than general-domain data and has already been used to train several open and closed models.
  • Llemma, the strongest open 7B and 34B base models for mathematical reasoning. These models are trained for up to 200B tokens primarily of OpenWebMath and show GPT-3.5-level performance with few-shot prompting even on held-out math evaluations.

selected publications

  1. llemma.png
    Llemma: An Open Language Model For Mathematics
    In International Conference on Learning Representations, 2024
  2. owm-color.png
    OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text
    Keiran Paster*Marco Dos Santos*Zhangir Azerbayev, and Jimmy Ba
    In International Conference on Learning Representations, 2024
  3. steve-1.gif
    STEVE-1: A Generative Model for Text-to-Behavior in Minecraft
    In Advances in Neural Information Processing Systems, 2023
    Spotlight
  4. ape-algo.gif
    Large Language Models are Human-Level Prompt Engineers
    In International Conference on Learning Representations, 2023
  5. esper.gif
    You Can’t Count on Luck: Why Decision Transformers and RvS Fail in Stochastic Environments
    Keiran PasterSheila A. McIlraith, and Jimmy Ba
    In Advances in Neural Information Processing Systems, 2022
  6. glamor.gif
    Planning from Pixels using Inverse Dynamics Models
    Keiran PasterSheila A. McIlraith, and Jimmy Ba
    In International Conference on Learning Representations, 2021