When
David Jurgens
Associate Professor
School of Information
Department of Computer Science & Engineering
University of Michigan
https://arizona.zoom.us/j/83216684767
Guiding LLM Reasoning: A Path to Robust and Human-Aligned Decisions
Abstract: This talk explores the reasoning processes of Large Language Models (LLMs), arguing that analyzing and shaping how a model arrives at a conclusion is key to building more robust and alignable systems. We demonstrate this principle in the high-stakes domain of moral reasoning, showing how models can be steered toward more human-aligned ethical problem-solving by providing more structure on how to reason with ethical principles. This approach of shaping reasoning also extends to general tasks, where we show that 'distilling' complex instructions into simpler, effective policies can improve a model's ability to generalize and solve novel problems. We then extend this analysis to a multilingual context, showing that models may spontaneously reason by code-switching (language mixing) within the reasoning trace. By characterizing these behaviors, we seek to better understand model performance in multilingual settings and how it may differ from human cognition. Further, we show that the language mixture of these traces can be strongly influenced by using prompts to steer models towards certain behaviors. This talk synthesizes multiple angles of reasoning to offer new perspectives on analyzing and guiding the problem-solving pathways of LLMs.