How LoRA Remembers? A Parametric Memory Law for LLM Finetuning

Ziwen Xu, Haiwen Hong, Linsong Yu, Benglei Cui, Longtao Huang, Hui Xue, Ningyu Zhang

arXiv:2605.30260 · 2026-05-29 공개 · arXiv · PDF

llm-finetuning lora token-level-analysis power-law threshold-optimization memory-fidelity parametric-memory loss-reduction

Abstract

Large Language Models (LLMs) must continuously learn and update knowledge to remain effective in dynamic real-world environments. While Low-Rank Adaptation (LoRA) is widely used for such memory updates, existing studies mainly rely on qualitative downstream evaluations, leaving the quantitative capacity limits and underlying dynamics of exact parametric memory largely unexplored. To bridge this gap, we employ LoRA as a controlled memory capacity probe within the latent space to systematically quantify exact parametric memory. We introduce the Parametric Memory Law, a robust power law linking loss reduction Delta L to effective parameters and sequence length. At the token level, fine-grained analysis reveals a deterministic phase transition, demonstrating that a prediction probability of p > 0.5 constitutes a sufficient condition for verbatim recall under greedy decoding. Driven by these insights, we introduce MemFT, a threshold-guided optimization strategy that dynamically redistributes the training budget toward sub-threshold tokens. Empirical evaluations demonstrate that MemFT can enhance memory fidelity and efficiency. Code will be released at https://github.com/zjunlp/ParametricMemoryLaw.

한국어 요약

📋 한 줄 요약

**[LoRA / Parametric Memory]** Parametric Memory Law — LoRA의 loss 감소(ΔL)가 effective parameter·sequence length에 멱법칙 따름, 토큰별 p>0.5가 verbatim recall 충분조건; MemFT가 sub-threshold 토큰에 budget 재분배.

🎯 핵심 기여도

💡 핵심 아이디어

LoRA의 parametric memory는 effective parameter·sequence length에 대한 멱법칙으로 정량화되며, 토큰 단위의 phase transition(p=0.5) 발견이 학습 budget을 sub-threshold 토큰에 재분배하는 명시적 최적화 전략을 가능하게 한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: LoRA 기반 지식 업데이트의 capacity·dynamics의 첫 정량 법칙 제시, 토큰 단위 phase transition 발견의 실용 가치(budget 재분배 직접 가이드), 멱법칙·threshold optimization의 결합이 ML scaling 연구와 정합. **한계**: Greedy decoding 가정·다른 sampling에서의 동작 추가 검증, LoRA 외 다른 PEFT 방법으로 일반화 후속, 매우 long sequence·구조화 지식의 멱법칙 검증 범위.

🚀 실용적 활용