Language models face a cold start problem when learning to reason...
The CDS Monthly Research Feature |
Language models face a cold start problem when learning to reason. If a model doesn’t already exhibit cognitive behaviors like verifying its own answers or backtracking from mistakes, no amount of reinforcement learning will teach it those skills.
CDS Associate Professor of Computer Science and Data Science Greg Durrett and his collaborators found a way around this obstacle. In their paper “SkillFactory: Self-Distillation for Learning Cognitive Behaviors,” they showed that models can learn complex reasoning behaviors from their own outputs, without a stronger model to teach it.
|
|
|
Within months of being born, babies achieve a stunning cognitive milestone: simply by observing the world around them, they learn to track objects and understand that things continue to exist even when they move out of sight. Replicating this natural learning process in artificial intelligence — specifically using only raw video footage without external help — has historically forced researchers to prioritize either object recognition or motion tracking, rarely achieving high performance in both simultaneously.
Now, new research from Assistant Professor of Computer Science and Data Science Mengye Ren and his colleagues at CDS proposes a solution that bridges this gap, offering a way for machines to learn from the world simply by observing it.
|
The leaderboard rankings that tech giants use to claim dominance in artificial intelligence are not nearly as robust as they’re commonly understood to be. New research from CDS PhD student Jingtong Su, CDS PhD alumnus Jianyu Zhang, and their collaborators reveals that changing a single, meaningless character in a test prompt can swing a model’s score by enough points to erase three years of progress.
|
Applications for the CDS MS programs for Fall 2026 are due February 14, 5:00pm ET. See some helpful links below:
|
|
|
Copyright (C) 2026 Center for Data Science. All rights reserved.
|
|
|
|
