Tuning Language Models by Mixture-of-Depths Ensemble
Authors: Haoyan Luo, Lucia Specia
Organizations: University of Cambridge · Imperial College London
Abstract
Transformer-based Large Language Models (LLMs) traditionally rely on final-layer loss for finetuning and final-layer representations for predictions, potentially overlooking the predictive power embedded in late layers. Interpretability tools such as the logit lens show that late-layer representations already carry largely formed, task-relevant predictions; here we ask whether that observation can be turned into an actionable training signal. We find that focusing tuning effort on these layers can yield losses comparable to those of the final layer, with complementary test-time behaviour. Building on this, we introduce a tuning framework, Mixture-of-Depths Ensemble (MoDE), which treats the late layers as an ensemble that contributes to the final logits through learned routing weights. MoDE can be applied on top of any existing tuning method (e.g., LoRA) and, in our experiments, modestly improves reasoning performance at a small parameter overhead. We present MoDE as a mechanism study showing that late-layer logits can be made directly useful for tuning, and that they can substitute for substantially larger trainable modules with comparable performance.
Parameter-efficient fine-tuning (PEFT) has become a practical route for adapting large language models to downstream tasks, with LoRA-style methods being particularly attractive because they are inexpensive to train and easy to deploy. Most LoRA variants, however, revise the update rule within the weight space of each layer and leave the intermediate representations formed by deeper layers largely unused. We propose Echo-LoRA, a cross-layer representation injection method for parameter-efficient fine-tuning. During training, Echo-LoRA collects boundary hidden states from deeper source layers, aggregates them into a sample-level echo representation, and uses lightweight projection and gating networks to inject the resulting signal into shallow LoRA or DoRA modules. Answer-only masking, masked distillation, and stochastic routing are used to keep this auxiliary path stable and to reduce the gap between training and inference. On eight commonsense reasoning benchmarks, Echo-LoRA exceeds the reported LoRA baselines by 5.7 percentage points on average across LLaMA-7B, LLaMA2-7B, and LLaMA3-8B. Under reproduced LoRA baselines in our unified implementation, the average gain is 3.0 points; when combined with DoRA, the gain is 2.7 points. The Echo path is discarded after training, so the deployed model keeps the original low-rank LoRA/DoRA form and adds neither inference-time parameters nor inference computation.
Full fine-tuning remains a strong way to adapt pretrained LLMs, but it updates all weights and can be expensive. LoRA reduces the number of trainable parameters, but it does not directly answer which pretrained components should be trained and which can be frozen during adaptation. We introduce TopoTuner, a topology-guided fine-tuning framework for selective freezing of attention projection matrices. \method treats each projection matrix as a row cloud and uses Wasserstein distances between persistence diagrams to measure how its topology changes during fine-tuning. TopoTuner learns a reusable freezing profile from a source dataset and transfers it to efficiently fine-tune models on out-of-domain datasets, evaluating whether task-specific topological drift generalizes across question answering and sentiment analysis tasks. Across LLaMA-3.1-8B, Mistral-7B-v0.3, and Qwen3-8B-Base, TopoTuner is competitive with full fine-tuning while training only 1-2% of the model parameters, and outperforms LoRA in 7 out of 9 model-dataset settings, which can change up to 39.57% of the projection parameters. Along with minimized updates, TopoTuner reduces training time by 20.4% relative to full fine-tuning and 5.5% relative to LoRA on average. TopoTuner opens a new direction for reusable freezing profiles, where fine-tuning behavior learned on one dataset can be shared across multiple tasks.
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× training speedup.