2 upvotes
New self-distillation trick boosts reasoning without any labeled data
What is Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models?
The paper introduces a method that deliberately weakens a large reasoning model during training, then uses those weaker outputs as a self-teacher to improve the original model. It works with only raw prompts, no external teacher or annotations needed.
Who it is for
Anyone building or fine-tuning reasoning agents can get stronger performance without costly data labeling or third-party models.