Understanding Knowledge Distillation in Post-Training: When It Helps and When It Fails

• URL: http://arxiv.org/abs/2606.22942v1
• Date: Mon, 22 Jun 2026 07:19:43 GMT
• Title: Understanding Knowledge Distillation in Post-Training: When It Helps and When It Fails
• Abstract summary: Large language models (LLMs) achieve strong performance across many tasks, but their high computational cost limits deployment in resource-constrained environments.<n> Knowledge Distillation (KD) offers a practical solution by transferring knowledge from a teacher model of a larger size to a smaller student model.
• Score: 16.73080036450313
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models (LLMs) achieve strong performance across many tasks, but their high computational cost limits deployment in resource-constrained environments. Knowledge Distillation (KD) offers a practical solution by transferring knowledge from a teacher model of a larger size to a smaller student model. While prior work has mainly examined task-specific or small-scale settings, the post-training stage for building general instruction-following models has received limited attention. In this paper, we conduct a systematic study of KD in post-training using the large-scale Tulu 3 dataset. We find that KD outperforms supervised fine-tuning (SFT) in low-data regimes, but its advantage diminishes as more training data is added. Distilling from a stronger instruction-tuned teacher restores substantial gains even with abundant data, indicating that KD remains effective when the teacher provides knowledge that the student cannot easily acquire from the training data alone. We further study domain-specific, low-resource scenarios and propose a two-stage KD strategy that leverages synthetic teacher-labeled data followed by refinement on human annotations. This method consistently improves student performance, providing practical guidance for building compact models in data-scarce environments.

Related papers

• Active Data Curation Effectively Distills Large-Scale Multimodal Models [66.23057263509027]
  Knowledge distillation (KD) is the de facto standard for compressing large-scale models into smaller ones.<n>In this work we explore an alternative, yet simple approach -- active data curation as effective distillation for contrastive multimodal pretraining.<n>Our simple online batch selection method, ACID, outperforms strong KD baselines across various model-, data- and compute-configurations.
  arXiv  Detail & Related papers  (2024-11-27T18:50:15Z)
• MiniPLM: Knowledge Distillation for Pre-Training Language Models [109.83741809808483]
  MiniPLM is a framework for pre-training student language models (LMs) using large teacher LMs.<n>For efficiency, MiniPLM performs offline teacher inference, allowing KD for multiple student LMs without adding training costs.<n>For flexibility, MiniPLM operates solely on the training corpus, enabling KD across model families.
  arXiv  Detail & Related papers  (2024-10-22T17:40:32Z)
• Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling [81.00825302340984]
  We introduce Speculative Knowledge Distillation (SKD) to generate high-quality training data on-the-fly.<n>In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution.<n>We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following.
  arXiv  Detail & Related papers  (2024-10-15T06:51:25Z)
• DistiLLM: Towards Streamlined Distillation for Large Language Models [53.46759297929675]
  DistiLLM is a more effective and efficient KD framework for auto-regressive language models. DisiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs.
  arXiv  Detail & Related papers  (2024-02-06T11:10:35Z)
• Robustness-Reinforced Knowledge Distillation with Correlation Distance and Network Pruning [3.1423836318272773]
  Knowledge distillation (KD) improves the performance of efficient and lightweight models.<n>Most existing KD techniques rely on Kullback-Leibler (KL) divergence.<n>We propose a Robustness-Reinforced Knowledge Distillation (R2KD) that leverages correlation distance and network pruning.
  arXiv  Detail & Related papers  (2023-11-23T11:34:48Z)
• Comparative Knowledge Distillation [102.35425896967791]
  Traditional Knowledge Distillation (KD) assumes readily available access to teacher models for frequent inference. We propose Comparative Knowledge Distillation (CKD), which encourages student models to understand the nuanced differences in a teacher model's interpretations of samples. CKD consistently outperforms state of the art data augmentation and KD techniques.
  arXiv  Detail & Related papers  (2023-11-03T21:55:33Z)
• Talking Models: Distill Pre-trained Knowledge to Downstream Models via Interactive Communication [25.653517213641575]
  We develop an interactive communication process to help students of downstream tasks learn effectively from pre-trained foundation models. Our design is inspired by the way humans learn from teachers who can explain knowledge in a way that meets the students' needs.
  arXiv  Detail & Related papers  (2023-10-04T22:22:21Z)
• Knowledge Distillation as Efficient Pre-training: Faster Convergence, Higher Data-efficiency, and Better Transferability [53.27240222619834]
  Knowledge Distillation as Efficient Pre-training aims to efficiently transfer the learned feature representation from pre-trained models to new student models for future downstream tasks. Our method performs comparably with supervised pre-training counterparts in 3 downstream tasks and 9 downstream datasets requiring 10x less data and 5x less pre-training time.
  arXiv  Detail & Related papers  (2022-03-10T06:23:41Z)
• Modality-specific Distillation [30.190082262375395]
  We propose modality-specific distillation (MSD) to effectively transfer knowledge from a teacher on multimodal datasets. Our idea aims at mimicking a teacher's modality-specific predictions by introducing an auxiliary loss term for each modality. Because each modality has different importance for predictions, we also propose weighting approaches for the auxiliary losses.
  arXiv  Detail & Related papers  (2021-01-06T05:45:07Z)
• MixKD: Towards Efficient Distillation of Large-scale Language Models [129.73786264834894]
  We propose MixKD, a data-agnostic distillation framework, to endow the resulting model with stronger generalization ability. We prove from a theoretical perspective that under reasonable conditions MixKD gives rise to a smaller gap between the error and the empirical error. Experiments under a limited-data setting and ablation studies further demonstrate the advantages of the proposed approach.
  arXiv  Detail & Related papers  (2020-11-01T18:47:51Z)

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.