Flexible Feature Distillation for Large Language Models

• URL: http://arxiv.org/abs/2507.10155v2
• Date: Thu, 02 Oct 2025 14:05:51 GMT
• Title: Flexible Feature Distillation for Large Language Models
• Authors: Khouloud Saadi, Di Wang,
• Abstract summary: We propose a parameter-free framework for task-driven feature distillation for large language models (LLMs)<n>Instead of projecting the entire teacher representation, Flex-KD uses gradient-based scores to identify the most task-relevant dimensions of the teacher's hidden states.<n>Experiments show that Flex-KD consistently boosts student performance, achieving up to a 3.75 percent performance gain over the linear projection baseline.
• Score: 4.277471273303775
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Knowledge distillation (KD) has become a cornerstone for compressing large language models (LLMs). However, existing LLM-KD methods have primarily focused on logit-based approaches, which achieve good performance but overlook the rich internal representations of LLMs. Feature-level KD could leverage this structure to provide complementary benefits, yet it remains underexplored because current feature-KD approaches typically assume identical teacher-student hidden sizes, a restrictive and unrealistic assumption. A common workaround is to train a linear projector to align their feature spaces; however, this introduces additional parameters, distorts teacher embeddings, and often degrades downstream performance, especially in generative tasks. We propose Flex-KD, a parameter-free framework for task-driven feature distillation for LLMs. Instead of projecting the entire teacher representation, Flex-KD uses gradient-based scores to identify the most task-relevant dimensions of the teacher's hidden states and distills only this subspace into the student. This ensures that the student's limited capacity is allocated to informative components, while avoiding projector-induced distortion and extra parameters. Flex-KD integrates seamlessly with existing KD pipelines and supports differing teacher-student hidden sizes. Extensive experiments across both classification and generative tasks, i.e., instruction-following and summarization, show that Flex-KD consistently boosts student performance, achieving up to a 3.75 percent performance gain over the linear projection baseline.

Related papers

• AdaSwitch: Adaptive Switching Generation for Knowledge Distillation [58.647880811071495]
  Small language models (SLMs) are crucial for applications with strict latency and computational constraints.<n>We propose AdaSwitch, a novel approach that combines on-policy and off-policy generation at the token level.<n>AdaSwitch consistently improves accuracy, offering a practical and effective method for distilling SLMs with acceptable additional overhead.
  arXiv  Detail & Related papers  (2025-10-09T06:38:37Z)
• Intra-class Patch Swap for Self-Distillation [3.282914142012984]
  We propose a teacher-free distillation framework based on a single student network.<n>Our approach is built on a simple yet highly effective augmentation, called intra-class patch swap augmentation.<n>Our method consistently outperforms both existing self-distillation baselines and conventional teacher-based KD approaches.
  arXiv  Detail & Related papers  (2025-05-20T09:30:19Z)
• Learning from Stochastic Teacher Representations Using Student-Guided Knowledge Distillation [64.15918654558816]
  Self-distillation (SSD) training strategy is introduced for filtering and weighting teacher representation to distill from task-relevant representations only.<n> Experimental results on real-world affective computing, wearable/biosignal datasets from the UCR Archive, the HAR dataset, and image classification datasets show that the proposed SSD method can outperform state-of-the-art methods.
  arXiv  Detail & Related papers  (2025-04-19T14:08:56Z)
• Fuse Before Transfer: Knowledge Fusion for Heterogeneous Distillation [52.0297393822012]
  We introduce an assistant model as a bridge to facilitate smooth feature knowledge transfer between heterogeneous teachers and students.<n>Within our proposed design principle, the assistant model combines the advantages of cross-architecture inductive biases and module functions.<n>Our proposed method is evaluated across some homogeneous model pairs and arbitrary heterogeneous combinations of CNNs, ViTs, spatial KDs.
  arXiv  Detail & Related papers  (2024-10-16T08:02:49Z)
• Linear Projections of Teacher Embeddings for Few-Class Distillation [14.99228980898161]
  Knowledge Distillation (KD) has emerged as a promising approach for transferring knowledge from a larger, more complex teacher model to a smaller student model. We introduce a novel method for distilling knowledge from the teacher's model representations, which we term Learning Embedding Linear Projections (LELP) Our experimental evaluation on large-scale NLP benchmarks like Amazon Reviews and Sentiment140 demonstrate the LELP is consistently competitive with, and typically superior to, existing state-of-the-art distillation algorithms for binary and few-class problems.
  arXiv  Detail & Related papers  (2024-09-30T16:07:34Z)
• Multi-Granularity Semantic Revision for Large Language Model Distillation [66.03746866578274]
  We propose a multi-granularity semantic revision method for LLM distillation. At the sequence level, we propose a sequence correction and re-generation strategy. At the token level, we design a distribution adaptive clipping Kullback-Leibler loss as the distillation objective function. At the span level, we leverage the span priors of a sequence to compute the probability correlations within spans, and constrain the teacher and student's probability correlations to be consistent.
  arXiv  Detail & Related papers  (2024-07-14T03:51:49Z)
• Fully Fine-tuned CLIP Models are Efficient Few-Shot Learners [8.707819647492467]
  We explore capturing the task-specific information via meticulous refinement of entire Vision-Language Models (VLMs) To mitigate these issues, we propose a framework named CLIP-CITE via designing a discriminative visual-text task.
  arXiv  Detail & Related papers  (2024-07-04T15:22:54Z)
• Direct Preference Knowledge Distillation for Large Language Models [73.50849692633953]
  We propose Direct Preference Knowledge Distillation (DPKD) for large language models (LLMs)<n>We re-formulate KD of LLMs into two stages: first optimizing and objective consisting of implicit reward and reverse KL divergence.<n>We prove the value and effectiveness of the introduced implicit reward and output preference in KD through experiments and theoretical analysis.
  arXiv  Detail & Related papers  (2024-06-28T09:23:40Z)
• PLaD: Preference-based Large Language Model Distillation with Pseudo-Preference Pairs [47.35598271306371]
  Large Language Models (LLMs) have exhibited impressive capabilities in various tasks, yet their vast parameter sizes restrict their applicability in resource-constrained settings. Knowledge distillation (KD) offers a viable solution by transferring expertise from large teacher models to compact student models. We present PLaD, a novel preference-based LLM distillation framework.
  arXiv  Detail & Related papers  (2024-06-05T03:08:25Z)
• Attention-guided Feature Distillation for Semantic Segmentation [8.344263189293578]
  This paper showcases the efficacy of a simple yet powerful method for utilizing refined feature maps to transfer attention.<n>The proposed Attention-guided Feature Distillation (AttnFD) method, employs the Convolutional Block Attention Module (CBAM)<n>It achieves state-of-the-art results in terms of improving the mean Intersection over Union (mIoU) of the student network on the PascalVoc 2012, Cityscapes, COCO, and CamVid datasets.
  arXiv  Detail & Related papers  (2024-03-08T16:57:47Z)
• RdimKD: Generic Distillation Paradigm by Dimensionality Reduction [16.977144350795488]
  Knowledge Distillation (KD) emerges as one of the most promising compression technologies to run advanced deep neural networks on resource-limited devices. In this work, we proposed an abstract and general paradigm for the KD task, referred to as DIMensionality Reduction KD (RdimKD) RdimKD solely relies on dimensionality reduction, with a very minor modification to naive L2 loss.
  arXiv  Detail & Related papers  (2023-12-14T07:34:08Z)
• Improving Knowledge Distillation via Regularizing Feature Norm and Direction [16.98806338782858]
  Knowledge distillation (KD) exploits a large well-trained model (i.e., teacher) to train a small student model on the same dataset for the same task. Treating teacher features as knowledge, prevailing methods of knowledge distillation train student by aligning its features with the teacher's, e.g., by minimizing the KL-divergence between their logits or L2 distance between their intermediate features. While it is natural to believe that better alignment of student features to the teacher better distills teacher knowledge, simply forcing this alignment does not directly contribute to the student's performance, e.g.
  arXiv  Detail & Related papers  (2023-05-26T15:05:19Z)
• Knowledge Diffusion for Distillation [53.908314960324915]
  The representation gap between teacher and student is an emerging topic in knowledge distillation (KD) We state that the essence of these methods is to discard the noisy information and distill the valuable information in the feature. We propose a novel KD method dubbed DiffKD, to explicitly denoise and match features using diffusion models.
  arXiv  Detail & Related papers  (2023-05-25T04:49:34Z)
• Efficient Feature Transformations for Discriminative and Generative Continual Learning [98.10425163678082]
  We propose a simple task-specific feature map transformation strategy for continual learning. Theses provide powerful flexibility for learning new tasks, achieved with minimal parameters added to the base architecture. We demonstrate the efficacy and efficiency of our method with an extensive set of experiments in discriminative (CIFAR-100 and ImageNet-1K) and generative sequences of tasks.
  arXiv  Detail & Related papers  (2021-03-25T01:48:14Z)
• Task-Feature Collaborative Learning with Application to Personalized Attribute Prediction [166.87111665908333]
  We propose a novel multi-task learning method called Task-Feature Collaborative Learning (TFCL) Specifically, we first propose a base model with a heterogeneous block-diagonal structure regularizer to leverage the collaborative grouping of features and tasks. As a practical extension, we extend the base model by allowing overlapping features and differentiating the hard tasks.
  arXiv  Detail & Related papers  (2020-04-29T02:32:04Z)

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.