Quantification of Large Language Model Distillation

TL;DR

This paper introduces a framework to systematically evaluate and quantify the extent of knowledge transfer and homogenization in large language model distillation, highlighting differences among models and emphasizing transparency for robustness.

Contribution

It proposes a novel method to measure distillation effects in LLMs, focusing on identity perception and response similarity, enhancing understanding of model homogenization.

Findings

High distillation degrees in most LLMs except Claude, Doubao, and Gemini.

Base LLMs exhibit higher distillation degrees than aligned LLMs.

The framework improves transparency and understanding of LLM data distillation processes.

Abstract

Model distillation is a fundamental technique in building large language models (LLMs), transferring knowledge from a teacher model to a student model. However, distillation can lead to model homogenization, reducing diversity among models and impairing their ability to robustly handle complex or novel tasks. These limitations underscore the need to systematically quantify the distillation process and its impact. In this work, we propose a framework to evaluate and quantify model distillation. Our method addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information, and (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. Experimental results demonstrate two key insights: (1) Well-known closed-source and open-source LLMs usually exhibit high distillation degrees, except for Claude, Doubao, and Gemini. (2) Base LLMs show higher distillation degrees compared to aligned LLMs. By offering a systematic approach to improve the transparency of LLM data distillation, we call for LLMs with more independent development and more transparent technical reports to improve LLMs' robustness and safety. The code and data are available under https://github.com/Aegis1863/LLMs-Distillation-Quantification.