Fine-Tuning & Adaptation · Efficient AI · Language Models

LoRA vs Full Fine-tuning Explained: The Intruder Dimensions Paper

MIT shows LoRA and full fine-tuning reach similar accuracy through different weights: LoRA adds intruder dimensions that carry most of its forgetting and can be scaled down after training at almost no accuracy cost.

LoRA vs Full Fine-tuning Explained: The Intruder Dimensions Paper

Quick answer

The paper asks whether LoRA and full fine-tuning learn the same solution when they reach the same accuracy, and answers no. Looking at the singular value decomposition of fine-tuned weight matrices, LoRA-trained matrices contain new singular vectors with large singular values that are nearly orthogonal to anything in the pretrained matrix. The authors call them intruder dimensions. Fully fine-tuned matrices do not have them; their top singular vectors stay close to the pretrained ones. The intruders matter because they carry the forgetting: shrinking their singular values after training cuts loss on the pretraining distribution by double-digit percentages while moving task accuracy by under a point in most settings. The paper is the reference for anyone who has read the LoRA claim that low-rank adaptation matches full fine-tuning and wants to know what that claim leaves out.

What an intruder dimension is

Take a pretrained weight matrix, fine-tune it, and compare the singular vectors before and after. Full fine-tuning on LLaMA2-7B leaves the similarity matrix close to the identity: the k-th singular vector after training points roughly where the k-th one pointed before. LoRA produces a different picture. Among the top singular vectors of the fine-tuned matrix there are new directions whose maximum cosine similarity to any pretrained singular vector is below a threshold (the paper sweeps it and uses 0.5 in the main text), and they sit at high rank, meaning large singular value. The detection procedure is Algorithm 1: for each of the top k singular vectors after fine-tuning, find the most similar pretrained singular vector; if that similarity is under the threshold, it is an intruder.

The paper shows the pattern on RoBERTa-base across six classification tasks and on LLaMA-7B and LLaMA2-7B fine-tuned on Alpaca, MetaMathQA and Magicoder-Evol-Instruct. Intruder dimensions appear at low LoRA ranks and shrink in number as rank increases; at r=64 and above LoRA solutions look more like full fine-tuning spectrally. Figure 5 tracks a single intruder over training: it appears early, its singular value grows, and it climbs the ranking as the run proceeds.

Key results

  • Same accuracy, different weights (Table 2, RoBERTa-base, alpha = 2r): MNLI full fine-tuning 0.8745, LoRA r=8 0.8704, r=16 0.8739; QQP full 0.9152, LoRA r=8 0.9093. The accuracy gap is under a point while the spectral structure differs qualitatively.
  • LoRA forgets less, even at matched fit (Figure 6): on LLaMA2-7B, measured on unrelated held-out tasks, and on RoBERTa via pretraining pseudo-loss, LoRA retains more of the pretraining distribution than full fine-tuning at the same downstream accuracy. The authors extend the earlier finding that LoRA forgets less by ruling out underfitting as the explanation.
  • But not always: with different choices of LoRA alpha and learning rate, LoRA runs with nearly identical task accuracy show starkly different generalization, including cases where LoRA forgets more than full fine-tuning. Runs with the fewest intruder dimensions generalize best.
  • Intruders cause the forgetting (Tables 3 and 4): scaling down the top intruder dimension in each matrix after training reduces pretraining loss far more than it reduces accuracy. RoBERTa QQP r=8: forgetting down 45.0% for a 0.6% accuracy drop; MNLI r=8: forgetting down 19.8% for 0.6%. LLaMA2-7B MetaMath r=16: forgetting down 28.2% for a 2.7% drop; Magicoder r=16: down 13.2% for 0.1%. Halving an intruder always costs under two percent accuracy on RoBERTa.
  • Continual learning (Figure 9): training sequentially on six tasks, full fine-tuning retains earlier-task performance better than LoRA because LoRA accumulates intruder dimensions across tasks.
  • What reduces intruders: higher rank, freezing a random orthogonal A and training only B (Figure 10), and lower learning rates or alpha (Appendix A.4).

Why the two methods diverge

The paper’s mechanistic account is the product parameterization. Full fine-tuning updates the whole matrix with a gradient that is, at the first step, a sum of rank-one terms aligned with the data. LoRA updates B and A separately, and because A is random at initialization, the update B A projects the gradient through a random low-dimensional subspace; the result is a direction that has no reason to align with the pretrained spectrum. Appendix A.2 shows that adding a random vector to the pretrained matrix reproduces intruder dimensions well, and A.3 derives the first-step difference between the two update rules. This is also why training only B on top of a fixed orthogonal A cuts intruders sharply.

Limits and open questions

The models are RoBERTa-base and 7B LLaMA variants; whether intruder dimensions behave the same in 70B-class models or under QLoRA-style quantization is not tested. Forgetting is measured by pretraining loss and by held-out task accuracy, both proxies for the capability loss users care about. The intervention of scaling down intruders is a post-hoc diagnostic, not a training recipe; the paper shows it works but does not propose a fixed schedule for choosing the scaling factor. And the study covers vanilla LoRA with a brief look at variants in Appendix O, so claims about DoRA or rank-stabilized variants should be checked against that appendix rather than assumed. The head-to-head with numbers from this paper and the original LoRA results is in LoRA vs Full Fine-tuning.

FAQ

What is the difference between LoRA and full fine-tuning according to this paper?

They reach similar task accuracy through different weight changes. LoRA introduces intruder dimensions, new high-ranking singular vectors nearly orthogonal to the pretrained spectrum, while full fine-tuning keeps the top singular vectors close to their pretrained directions.

Does LoRA forget less than full fine-tuning?

Usually, and the paper confirms it holds even when task accuracy is matched. But it depends on hyperparameters: with some alpha and learning-rate settings LoRA forgets more than full fine-tuning, and those runs have more intruder dimensions.

What happens if you scale down the intruder dimensions after LoRA fine-tuning?

Forgetting drops sharply while accuracy barely moves. On RoBERTa QQP with r=8, pretraining loss forgetting fell 45% for a 0.6% accuracy drop; on LLaMA2-7B MetaMath with r=16 it fell 28.2% for a 2.7% drop.

Which LoRA settings produce fewer intruder dimensions?

Higher rank (r=64 and above behaves more like full fine-tuning), training only the B matrix over a fixed orthogonal A, and lower learning rates or alpha. Continual fine-tuning across many tasks does the opposite and accumulates them.