Researchers · Interviewed at COLM 2026
Stanford University

COLM 2026 · October 20, 2026 · 4 min
You ask an LLM "Did I mess up?" and want an honest answer. The model assumes you want reassurance and tells you that you’re right. New work presented at COLM 2026 shows that this hidden assumption about the user drives sycophancy, and that you can steer it.
In today's COLM 2026 poster highlight, Myra Cheng, a final-year PhD candidate in Computer Science at Stanford, presents her work on verbalizing assumptions to understand and control LM sycophancy.