Symmetry-based framework for analyzing LLM preference conflicts
July 31, 2026
A new symmetry-based framework measures how LLMs resolve conflicts between competing specifications, such as natural language versus formal language. Testing on 550 mathematical conflict instances reveals how model choices shift across different representation types.
HOW THIS AFFECTS YOU
●
researcherYou can systematically analyze and attribute how models prioritize conflicting instructions during reasoning tasks.