Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs

• URL: http://arxiv.org/abs/2606.05516v1
• Date: Wed, 03 Jun 2026 23:42:09 GMT
• Title: Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs
• Abstract summary: Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models (LLMs) using only forward passes.<n>In this work, we reveal a surprising phenomenon: ZO fine-tuning is sharply dominated by a single decoding layer.<n>Experiments on LLaMA2-7B and Qwen3-8B across nine benchmarks show that dominant-layer ZO fine-tuning improves average performance over full-model MeZO and LoRA-based ZO fine-tuning.
• Score: 32.78500862748985
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models (LLMs) using only forward passes, but it remains unclear how useful adaptation is distributed across layers. In this work, we reveal a surprising phenomenon: ZO fine-tuning is sharply dominated by a single decoding layer. Across multiple LLM families and downstream tasks, fine-tuning this dominant layer alone consistently matches or even exceeds full-model ZO fine-tuning. We further show that the dominant layer is task-agnostic but model-specific, and can be identified before training through a simple inference-only analysis of activation outliers. Specifically, the dominant layer consistently aligns with the first activation-outlier layer in the pre-trained model. To explain this phenomenon, we analyze how perturbation effects propagate under ZO optimization. We find that the dominant layer combines two key properties: high perturbation sensitivity and early placement in the residual stream, allowing perturbation-induced effects to propagate and accumulate through remaining subsequent decoding layers. As a result, this layer produces disproportionately strong and stable optimization signals under forward-only updates. Extensive experiments on LLaMA2-7B and Qwen3-8B across nine benchmarks show that dominant-layer ZO fine-tuning improves average performance over full-model MeZO and LoRA-based ZO fine-tuning while achieving up to 4.52$\times$ training speedup.

Related papers

• Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding [92.1521161575198]
  We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer.<n> Experiments across dense and Mixture-of-Experts LLMs demonstrate consistent gains on challenging reasoning benchmarks.
  arXiv  Detail & Related papers  (2026-06-20T07:03:26Z)
• Layer Collapse in Diffusion Language Models [54.880703002010144]
  Diffusion language models (DLMs) have emerged as competitive alternatives to autoregressive (AR) language models.<n>We show that layer collapse in DLMs is not driven by undertraining but by overtraining.<n>Our findings have strong practical implications.
  arXiv  Detail & Related papers  (2026-05-07T14:39:40Z)
• LayerBoost: Layer-Aware Attention Reduction for Efficient LLMs [3.80555579179805]
  LayerBoost reduces inference latency and improves throughput by up to 68% at high.<n>It matches base model performance on several benchmarks, exhibits only minor degradations on others, and significantly outperforms state-of-the-art attention linearization methods.
  arXiv  Detail & Related papers  (2026-04-23T20:12:19Z)
• Layer by layer, module by module: Choose both for optimal OOD probing of ViT [16.482899285404145]
  We study the behavior of intermediate layers in pretrained vision transformers.<n>We find that distribution shift between pretraining and downstream data is the primary cause of performance degradation.
  arXiv  Detail & Related papers  (2026-03-05T15:23:41Z)
• GradPruner: Gradient-Guided Layer Pruning Enabling Efficient Fine-Tuning and Inference for LLMs [10.61152477422108]
  GradPruner can prune layers of Large Language Models guided by gradients in the early stages of fine-tuning.<n>Results demonstrate that GradPruner has achieved a parameter reduction of 40% with only a 0.99% decrease in accuracy.
  arXiv  Detail & Related papers  (2026-01-27T11:41:26Z)
• Determining Layer-wise Sparsity for Large Language Models Through a Theoretical Perspective [55.90119819642064]
  We address the challenge of determining the layer-wise sparsity rates of large language models (LLMs) through a theoretical perspective.<n>This refers to the cumulative effect of reconstruction errors throughout the sparsification process.<n>We derive a simple yet effective approach to layer-wise sparsity allocation that mitigates this issue.
  arXiv  Detail & Related papers  (2025-02-20T17:51:10Z)
• The Curse of Depth in Large Language Models [28.37870372690079]
  In large language models, nearly half of the layers are less effective than expected.<n>LayerNorm Scaling (LNS) scales the variance of output of the layer normalization inversely by the square root of its depth.<n>LNS consistently outperforms previous normalization and scaling techniques in enhancing LLM pre-training performance.
  arXiv  Detail & Related papers  (2025-02-09T07:03:36Z)
• Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning [65.23593936798662]
  We show that fine-tuning with LLM-generated data improves target task performance and reduces non-target task degradation.<n>This is the first work to provide an empirical explanation based on token perplexity reduction to mitigate catastrophic forgetting in LLMs after fine-tuning.
  arXiv  Detail & Related papers  (2025-01-24T08:18:56Z)
• Understanding Layer Significance in LLM Alignment [23.582520695083588]
  We propose identifying which layers within large language models are most critical to the alignment process.<n> Experimental results reveal that, despite substantial differences in alignment datasets, the important layers of a model exhibit nearly 90% overlap.<n>The results also indicate that freezing non-essential layers improves overall model performance, while selectively tuning the most critical layers significantly enhances fine-tuning efficiency with minimal performance loss.
  arXiv  Detail & Related papers  (2024-10-23T13:47:05Z)
• Zeroth-Order Fine-Tuning of LLMs in Random Subspaces [63.10833446782114]
  As language models grow in size, memory demands for backpropagation increase.<n>Zeroth-order (ZO) optimization methods offer a memory-efficient alternative.<n>In this paper, we propose Subspace Zero-order optimization to address the challenges posed by posed by high dimensionality perturbations.
  arXiv  Detail & Related papers  (2024-10-11T17:01:43Z)
• Model-Agnostic Multitask Fine-tuning for Few-shot Vision-Language Transfer Learning [59.38343286807997]
  We propose Model-Agnostic Multitask Fine-tuning (MAMF) for vision-language models on unseen tasks. Compared with model-agnostic meta-learning (MAML), MAMF discards the bi-level optimization and uses only first-order gradients. We show that MAMF consistently outperforms the classical fine-tuning method for few-shot transfer learning on five benchmark datasets.
  arXiv  Detail & Related papers  (2022-03-09T17:26:53Z)
• Layer-wise Conditioning Analysis in Exploring the Learning Dynamics of DNNs [115.35745188028169]
  We extend conditioning analysis to deep neural networks (DNNs) in order to investigate their learning dynamics. We show that batch normalization (BN) can stabilize the training, but sometimes result in the false impression of a local minimum. We experimentally observe that BN can improve the layer-wise conditioning of the optimization problem.
  arXiv  Detail & Related papers  (2020-02-25T11:40:27Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.