Machine Learning Street Talk (MLST) What’s the Magic Word? A Control Theory of LLM Prompting.
119 snips
Jun 5, 2024 
Aman Bhargava, a PhD student at Caltech, and Cameron Witkowski, a graduate student at the University of Toronto, dive into their groundbreaking research on controlling language models using control theory. They discuss how language models operate as discrete systems and the surprising impact of prompt engineering on outputs. By examining the "reachable set" of outputs, they reveal that even minor tweaks in prompts can lead to significant changes in generated text. Their insights could pave the way for more reliable and capable AI systems.
AI Snips
Chapters
Books
Transcript
Episode notes
Control Theory Perspective
Framing LLMs as discrete stochastic dynamical systems allows for a structured way to analyze their output behaviors. This perspective enhances understanding of controllability and the reachability of outputs.
Explore Control Inputs
Explore how different prompts influence the output of LLMs. Experiment with varying lengths and types of control inputs to discover the optimal ways to steer the models.
Adversarial Properties of LLMs
Like other systems, LLMs exhibit adversarial properties, showing that small input changes can influence their responses. The chaotic nature of some prompts reveals underlying complexities within LLM behavior.
