Researchers · Interviewed at ICLR 2026
Meta

ICLR 2026 · June 1, 2026 · 4 min
Current LLM alignment methods assume we can learn human preferences from preference datasets. But what if those datasets are fundamentally missing the diversity of human values?
At #ICLR 2026, Meta researcher Smitha Milli presents the Community Alignment Dataset, a new open-source resource designed to capture how preferences vary across people, cultures, and communities.