Wolfgang Lehrach | Code World Models for General Game Playing
While Large Language Models (LLMs) show promise in many domains, relying on them for direct policy generation in games often results in illegal moves and poor strategic play.
Get the latest news, advances in research, policy work, and education program updates from HAI in your inbox weekly.
Sign Up For Latest News
While Large Language Models (LLMs) show promise in many domains, relying on them for direct policy generation in games often results in illegal moves and poor strategic play.
The next, maybe final, frontier of AI research is automating the process of knowledge discovery. By pointing AI first at itself and then at the R&D process of all scientific domains, we can far exceed where human-led research can take us. Richard Socher argues the current pace of scientific discovery has slowed significantly over the past several decades as the fields of science fragment and the number of researchers focused on any one continues to shrink.

The next, maybe final, frontier of AI research is automating the process of knowledge discovery. By pointing AI first at itself and then at the R&D process of all scientific domains, we can far exceed where human-led research can take us. Richard Socher argues the current pace of scientific discovery has slowed significantly over the past several decades as the fields of science fragment and the number of researchers focused on any one continues to shrink.
This talk will explore the latest advances in computational protein design and structure prediction, highlighting NVIDIA technologies such as cuEquivariance and the Proteína family of models, including the latest RFdiffusion4 (RFD4)–Proteína developments. We will discuss how accelerated geometric deep learning and generative AI are enabling scalable, high-performance protein modeling and the design of novel proteins with tailored structures and functions.
Bio:
Youssef Nashed is a Senior Solutions Architect on NVIDIA's Healthcare and Life Sciences team. He joined NVIDIA from Bristol Myers Squibb (BMS), where he was a Senior Principal Scientist focused on machine learning and computational protein design. At BMS, Youssef led a team developing and deploying AI models for antibody discovery and development. He previously served as an Assistant Professor at Northwestern University and conducted research in structural biology at SLAC National Accelerator Laboratory at Stanford. Youssef began his career in high-performance and large-scale GPU computing at Argonne National Laboratory. His background spans computer science, mathematics, structural biology, and AI-driven protein engineering.
Location: CoDa W401

This talk will explore the latest advances in computational protein design and structure prediction, highlighting NVIDIA technologies such as cuEquivariance and the Proteína family of models, including the latest RFdiffusion4 (RFD4)–Proteína developments. We will discuss how accelerated geometric deep learning and generative AI are enabling scalable, high-performance protein modeling and the design of novel proteins with tailored structures and functions.
Bio:
Youssef Nashed is a Senior Solutions Architect on NVIDIA's Healthcare and Life Sciences team. He joined NVIDIA from Bristol Myers Squibb (BMS), where he was a Senior Principal Scientist focused on machine learning and computational protein design. At BMS, Youssef led a team developing and deploying AI models for antibody discovery and development. He previously served as an Assistant Professor at Northwestern University and conducted research in structural biology at SLAC National Accelerator Laboratory at Stanford. Youssef began his career in high-performance and large-scale GPU computing at Argonne National Laboratory. His background spans computer science, mathematics, structural biology, and AI-driven protein engineering.
Location: CoDa W401
In this talk, I present an approach that moves away from direct prompting, instead using LLMs as program synthesizers to bridge the gap between natural language rules and symbolic world models. The LLM receives a game description and example trajectories, and outputs an executable, symbolic world model (CWM) represented in Python. The trajectories also ensure the rules are correctly captured and aid in refining the CWM if they are not. Note that even trajectories containing only a single player's observations and actions can be used to help validate and refine CWMs. Furthermore, partially observed trajectories also allow comparisons between CWMs via a bound on the likelihood.
Given a CWM, Monte Carlo Tree Search (MCTS) or Reinforcement Learning (RL) methods can play the game, and gameplay can be further enhanced by adding in LLM-derived synthesized value functions. Imperfect information games are handled by having the LLM synthesize inference functions to impute information sets, or by directly training reinforcement learning policies on top of the CWM.