| Challenge: | Existing knowledge distillation methods for autoregressive large language models (LLMs) are not effective for reducing generation quality, but they can be useful for real-time applications. |
| Approach: | They propose a self-distillation framework that allows for effective KD without external teacher . they propose to use two modes of knowledge distillation to determine distillation direction . |
| Outcome: | The proposed framework outperforms existing methods with external teachers on instruction-following tasks. |
Similar Papers
Self-Evolution Knowledge Distillation for LLM-based Machine Translation (2025.coling-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation strategies for large language models minimize output distributions between student and teacher models indiscriminately for each token. |
| Approach: | They propose a distillation strategy that integrates teacher and one-hot distribution of ground truth into the student distribution as prior knowledge, which promotes the distillation process. |
| Outcome: | The proposed method brings an average improvement of approximately 1.4 SacreBLEU points across four translation directions in the WMT22 test sets. |
Knowledge Distillation for Language Models (2025.naacl-tutorial)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) aims to transfer knowledge from a teacher to a student . this tutorial will cover topics ranging from LLM sequence compression to LLM self-distillation . |
| Approach: | They propose to introduce intermediate-layer matching and prediction matching . they will then present advanced techniques such as reinforcement learning-based KD and multi-teacher distillation . |
| Outcome: | This tutorial aims to provide participants with a comprehensive understanding of the techniques and applications of knowledge distillation for language models. |
On-Policy Self-Distillation for Efficient Diffusion Language Models with Early-Stage Calibration (2026.findings-acl)
Copied to clipboard
Huaisheng Zhu, MingYu Liu, Junze Liu, Zhen Ge, Tian Wang, Jiri Gesi, Dakuo Wang, Weiqi Zhang, Houyu Zhang, Yufan Guo, Xian Li, Bing Yin, Sujay Sanghavi
| Challenge: | Recent studies have demonstrated that masked diffusion models (MDMs) can surpass autoregressive models (ARMs) in various tasks. |
| Approach: | They propose a method to calibrate early token predictions without demonstration data by distilling an unnormalized target distribution into the original model. |
| Outcome: | Experiments on math, planning, and RLHF tasks show that COPSD improves both effectiveness and efficiency, and further enhances performance when combined with supervised fine-tuning. |
Revisiting Knowledge Distillation for Autoregressive Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Autoregressive language models (LMs) are expensive and memory intensive, preventing the development of industrial applications. |
| Approach: | They propose an adaptive teaching approach to improve the KD of autoregressive language models by distilling knowledge into a small student model. |
| Outcome: | The proposed method can achieve consistent and significant performance gains across all model types and sizes. |
Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning (2024.acl-long)
Copied to clipboard
| Challenge: | Experimental results show that fine-tuning of large language models for specific tasks can be challenging . distribution shift during fine-timing can lead to performance degradation in general task capabilities . |
| Approach: | They propose a new approach that bridges the distribution gap between task datasets and LLMs by guiding fine-tuning with a distilled dataset generated by the model itself. |
| Outcome: | The proposed approach achieves comparable or superior performance on downstream tasks compared to the vanilla approach. |
A Self-Distillation Recipe for Neural Machine Translation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for Neural Machine Translation (NMT) have been proven effective in improving the performance of computer vision tasks without pre-training a teacher. |
| Approach: | They propose a rank-order augmented Pearson correlation loss and an iterative distillation method to prevent the discrepancy of predictions between the student and a stronger teacher from disturbing the training. |
| Outcome: | The proposed method can lead to significant improvements over the strong Transformer baseline on low/middle/high-resource tasks, obtaining comparable or better performance with fewer layers. |
Streamlining LLMs: Adaptive Knowledge Distillation for Tailored Language Models (2025.naacl-srw)
Copied to clipboard
| Challenge: | Large language models (LLMs) have transformative potential across industries, e.g., enhancing customer service, revolutionizing medical diagnostics, or identifying crises in news articles. |
| Approach: | They propose to distill compact, parameter-efficient tailored language models from LLMs for domain-specific tasks with comparable performance. |
| Outcome: | The proposed framework outperforms knowledge distillation frameworks in the crisis domain, where labeled data is scarce. |
Beyond One-Step Distillation: Bridging the Capacity Gap in Small Language Models via Multi-Step Knowledge Transfer (2026.eacl-srw)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) excel across diverse tasks but remain too large for efficient on-device deployment. |
| Approach: | They revisit multi-step knowledge distillation as an effective remedy . they demonstrate that MSKD improves ROUGE-L and perplexity over single-step approaches . |
| Outcome: | The proposed approach improves ROUGE-L and perplexity over single-step approaches . large language models are too large for efficient on-device deployment, the authors show . |
AD-KD: Attribution-Driven Knowledge Distillation for Language Model Compression (2023.acl-long)
Copied to clipboard
| Challenge: | Existing knowledge distillation methods focus on the transfer of model-specific knowledge but overlook data-specific information. |
| Approach: | They propose an attribution-driven knowledge distillation approach which explores the token-level rationale behind the teacher model and transfers attribution knowledge to the student model. |
| Outcome: | The proposed method outperforms state-of-the-art methods on the GLUE benchmark and shows that it is more efficient than existing methods. |
ToDi: Token-wise Distillation via Fine-Grained Divergence Control (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) offer impressive performance but are impractical for resource-constrained deployment due to high latency and energy consumption. |
| Approach: | They propose a method that adaptively combines FKL and RKL per token using a sigmoid-based weighting function derived from the teacher-student probability log-ratio. |
| Outcome: | The proposed method outperforms baselines using uniform or less granular strategies across instruction-following benchmarks. |