Reproduce “Updating Parametric Knowledge with Context Distillation Retains Post-Training Capabilities” on OpenResearch

About this paper

Shankar Padmanabhan, Mustafa Omer Gul, Tanya Goyal

Researchers at Cornell University developed Distillation via Split Contexts (DiSC), a method to update post-trained Large Language Models (LLMs) with new knowledge from documents while preserving their general capabilities. DiSC achieves superior knowledge adaptation with minimal forgetting, often outperforming standard fine-tuning by 5-12 points in domain adaptation and incurring approximately half the forgetting on instruction-following tasks.

Key insights

• Standard fine-tuning with a next-token prediction objective leads to substantial catastrophic forgetting, particularly of complex "post-training skills" like instruction-following, math, and coding, often reverting the model towards its pre-trained state.
• Using KL divergence as the primary learning objective in DiSC is more effective for balancing new knowledge internalization and preserving existing capabilities than auxiliary regularizers in fine-tuning.
• The extent of performance gain from pre-training to post-training correlates strongly with the subsequent performance drop due to fine-tuning, indicating that recently acquired, high-value capabilities are most vulnerable.

Summary by alphaXiv.