AI BriefingSelf-Distillation Enables Continual Learning
Key point
Self-Distillation Fine-Tuning reduced forgetting in demonstration-based learning.
Details
The authors propose Self-Distillation Fine-Tuning (SDFT), which directly generates on-policy signals in demonstration-based learning and causes less loss of prior abilities.
The core idea is to use the demonstration-conditioned model as its own teacher, performing self-distillation that leverages in-context learning. This mitigates the off-policy limitation inherent in existing supervised fine-tuning (SFT).
In experiments:
- On skill learning and knowledge acquisition tasks, it showed higher new-task accuracy than SFT.
- At the same time, it significantly reduced catastrophic forgetting.
- In sequential learning settings, a single model accumulated multiple skills over time while maintaining performance without degradation.
The authors conclude that these results demonstrate the potential of on-policy distillation as a practical path to achieving continual learning in demonstration-based learning.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.