FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning

Yequan Zhao, Ruijie Zhang, Liyan Tan, Niall Moran, Tong Qin, Zheng Zhang

arXiv:2605.22869 · 2026-05-25 공개 · arXiv · PDF

reinforcement-learning llama qlora llm-fine-tuning parameter-efficient-fine-tuning singular-value-decomposition visual-instruction-tuning spectral-preconditioning

Abstract

Both full fine-tuning (Full FT) and parameter-efficient fine-tuning methods such as LoRA introduce weight updates without accounting for the spectral structure established during pretraining. As a result, noisy gradients from limited fine-tuning data can perturb robust pretrained features. We identify spectral preconditioning as the missing ingredient: reparameterizing each weight matrix through its full-rank singular value decomposition (SVD) and freezing one singular basis constrains updates to the pretrained column space, yielding a preconditioned optimization scheme that outperforms unconstrained Full FT at the same trainable parameter count. Building on this insight, we propose FuRA (Full-Rank Adaptation), an efficient full-rank adaptation framework based on a block tensor-train factorization W = LSR, where the large core L is fixed to the pretrained block-wise SVD basis, while only the compact core R and the block-wise singular values S are optimized. This design simultaneously provides full-rank spectral preconditioning, preserves full-rank update expressivity, and achieves parameter, memory, and step-time efficiency comparable to LoRA. FuRA consistently outperforms Full FT across multiple settings, including LLM fine-tuning (+1.37 on LLaMA-3-8B commonsense reasoning), LLM reinforcement learning for mathematical reasoning, and visual instruction tuning for VLMs. Furthermore, the 4-bit quantized variant, QFuRA, also surpasses QLoRA. Code is available at https://github.com/olokevin/FuRA-NIPS

한국어 요약

📋 한 줄 요약

**[Spectral PEFT]** FuRA가 block-wise SVD 기반 W=LSR 분해로 L을 동결·S·R만 학습 — full-rank spectral preconditioning, LoRA 수준 파라미터·메모리·step 효율, LLaMA-3-8B commonsense +1.37·QFuRA가 QLoRA 능가.

🎯 핵심 기여도

💡 핵심 아이디어

Fine-tuning 효과는 trainable parameter 수가 아니라 spectral 구조의 보존·활용에 달려있으며, block tensor-train W=LSR로 pretrained SVD basis를 동결한 채 singular value와 compact core만 학습하면 full-rank expressivity와 LoRA 수준 효율을 동시 달성한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: PEFT의 spectral 구조 보존이라는 missing principle 정립, full-rank expressivity와 LoRA 효율의 동시 확보로 PEFT 패러다임 진전, 양자화까지 확장. **한계**: Block-wise SVD 사전 계산 비용, 매우 큰 모델에서 block 분할 hyperparameter 튜닝, 일부 도메인의 효과 검증 추가 필요.

🚀 실용적 활용