Pretraining Induces a Reusable Spectral Basis for Downstream Task Adaptation
Authors: Junjie Yu, Yue Wang, Zihan Deng, Yan Zhu, Wenxiao Ma, Quanying Liu
Abstract
Finetuning pretrained models occurs in a low-dimensional subspace of the full parameter space. Prior work has focused on characterizing this optimization subspace, but largely ignored the complementary question: why do certain directions remain unexplored during finetuning? Are these stable directions irrelevant to downstream tasks, or do they already encode task-relevant structure that requires no further adjustment? Answering this question is central to understanding how pretrained knowledge transfers. Through systematic spectral analysis across vision and language models, we show that the leading singular vectors of pretrained weight matrices remain highly stable under finetuning and are shared across unrelated downstream tasks, revealing that pretraining establishes a reusable spectral coordinate system. Models pretrained on larger datasets exhibit greater spectral stability under distribution shift or task change, directly linking pretraining scale to geometric transferability. Motivated by these findings, we propose a parameter-efficient method that freezes pretrained singular vectors and optimizes only leading spectral coefficients, achieving competitive performance on GLUE with 0.2% trainable parameters. Our results reveal that the stable directions encode transferable structure rather than irrelevant noise: successful pretraining discovers spectral bases that downstream tasks inherit and operate within.
Parameter-efficient finetuning (PEFT) has become the standard approach for adapting large language models, yet evaluations largely emphasize downstream accuracy while overlooking the retention of pretrained capabilities. We argue that PEFT should be assessed through the stability-plasticity dilemma: the trade-off between target-task adaptation and resistance to forgetting. We introduce PEFT-Arena, a benchmark that jointly measures downstream performance and general capability retention. Across methods, we find distinct stability-plasticity profiles; under comparable parameter budgets, orthogonal finetuning achieves the most favorable Pareto frontier. To explain these differences, we analyze PEFT updates from two geometric perspectives. In weight space, spectral analysis reveals how parameterizations interact with the pretrained singular-value structure. In activation space, retention metrics show whether finetuning preserves or distorts general-capability representations, with forgetting linked to non-isometric representation distortion. Finally, an analysis shows that final SFT checkpoints often overshoot a better target-retention operating point. Inspired by this, we present case studies of a post-hoc improvement with path-wise rewinding.
Fine-tuning large-scale pre-trained models is a recent prevalent paradigm for adapting general representations to specialized tasks. However, when a new version of a pre-trained model becomes available, expertise acquired through fine-tuning cannot be directly reused because it is tied to the parameterization of the original model, requiring another costly fine-tuning. To address this inefficiency, recent work uses task vectors, defined as the parameter difference between a fine-tuned model and its base model, to transfer expertise across models. While existing methods bridge disparate models by matching activations or gradients, a significant performance gap remains relative to direct fine-tuning, suggesting that these partial correspondences are insufficient. In this work, instead of viewing a task vector merely as a parameter offset, we revisit the formation of task vectors and show that they can be derived as accumulated bilinear interactions between input-side activations and output-side gradients. Motivated by this observation, we formulate task-vector transfer as a dual-space alignment problem and propose BiCo, a training-free framework for transferring task vectors through Bilinear Coordinate alignment. BiCo estimates orthogonal Procrustes mappings in both spaces using a single forward-backward pass on a small calibration set, without any parameter update. Across extensive computer vision and natural language processing benchmarks, BiCo consistently outperforms existing transfer methods across models that differ in width, depth, and pre-training configuration.
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