New self-distillation method recovers LLM performance by aligning internal manifolds

Researchers have introduced a novel framework called Self-Distillation Fine-Tuning (SDFT) designed to recover performance in large language models (LLMs) that have degraded due to factors like catastrophic forgetting during supervised fine-tuning, quantization, or pruning. The framework is supported by a theoretical explanation that posits LLM generative capability is tied to the high-dimensional manifold constructed by its hidden layers. Experiments using Centered Kernel Alignment (CKA) demonstrated a strong correlation between performance recovery and manifold alignment, suggesting self-distillation effectively aligns the student model's manifold with the teacher's optimal structure. AI

IMPACT This research offers a theoretical and practical method to counteract performance degradation in LLMs, potentially improving their reliability and longevity after fine-tuning.

RANK_REASON The cluster contains an academic paper detailing a new method for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Chi Liu
• large-language models
• Self-Distillation Fine-Tuning
• supervised fine-tuning

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →