Rethinking Kullback-Leibler Divergence in Knowledge Distillation for Large Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have been used in Knowledge Distillation (KD) to compress large models. |
| Approach: | They propose a Kullback-Leiber divergence method which adaptively allocates weights to combine RKL and FKL to reduce the size of Large Language Models (LLMs). |
| Outcome: | The proposed method outperforms baselines and improves diversity and quality of generated responses. |
Similar Papers
Balancing Forget Quality and Model Utility: A Reverse KL-Divergence Knowledge Distillation Approach for Better Unlearning in LLMs (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods for unlearning large language models struggle with forget quality and model utility, leading to over-unlearning or partial unlearning. |
| Approach: | They propose a method that uses reverse KL-divergence based knowledge distillation for unlearning to achieve significant forget quality while maintaining model utility. |
| Outcome: | The proposed method outperforms existing methods in forget quality and model utility with larger unlearning datasets. |
Sinkhorn Distance Minimization for Knowledge Distillation (2024.lrec-main)
Copied to clipboard
Xiao Cui, Yulei Qin, Yuting Gao, Enwei Zhang, Zihan Xu, Tong Wu, Ke Li, Xing Sun, Wengang Zhou, Houqiang Li
| Challenge: | Existing knowledge distillation methods investigate divergence measures but fail to deliver effective supervision when few distribution overlap exists between teacher and student. |
| Approach: | They propose a knowledge distillation method that exploits the Sinkhorn distance to ensure a nuanced assessment of the disparity between teacher and student distributions. |
| Outcome: | The proposed method outperforms state-of-the-art methods on all kinds of LLMs with encoder-only, encoder decoder, and decoded architectures. |
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. |
Efficient KL Divergence Estimation via Truncated Top-K Integration for Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for estimating KL divergence using only top-k tokens suffer from high variance or systematic bias. |
| Approach: | They propose a top-k Importance-weighted KL Estimator that exploits the Zipfian structure of language model distributions by integrating only the top-K tokens. |
| Outcome: | The proposed estimator outperforms existing estimators on multiple benchmarks while exhibiting lower variance. |
Establishing a Scale for Kullback-Leibler Divergence in Language Models Across Various Settings (2026.findings-acl)
Copied to clipboard
| Challenge: | Fig. 1 and 2 shows that log-likelihood vectors provide a consistent representation for language models . weight permutation symmetries and architectural dependencies hinder direct comparisons between models with different learning methods or designs. |
| Approach: | They propose a log-likelihood vector for comparing language models as probability distributions . they establish a consistent scale for KL divergence across various settings . |
| Outcome: | The proposed model comparisons show that the log-likelihood space is smaller than the weight space . the proposed model compares language models across checkpoints, model sizes, quantization, fine-tuning, and layers . |
Multi-Stage Balanced Distillation: Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a promising solution for large language models, but their deployment remains computationally expensive. |
| Approach: | They propose a framework which iteratively balances training data within a fixed computational budget and enables the transfer of knowledge from expensive teacher LLMs to smaller student models. |
| Outcome: | The proposed framework achieves state-of-the-art performance across diverse long-tailed datasets, enhancing both the efficiency and efficacy of the distilled models. |
LLMR: Knowledge Distillation with a Large Language Model-Induced Reward (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models have demonstrated remarkable performance in various NLP tasks, but are typically computationally expensive and difficult to be deployed in resource-constrained environments. |
| Approach: | They propose a knowledge distillation method based on a reward function induced from large language models. |
| Outcome: | The proposed method outperforms traditional methods on multiple datasets and tasks. |
f-Divergence Minimization for Sequence-Level Knowledge Distillation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing knowledge distillation approaches focus on minimizing a generalized f-divergence function. |
| Approach: | They propose a framework which formulates sequence-level knowledge distillation as minimizing a generalized f-divergence function. |
| Outcome: | The proposed framework outperforms existing methods and reduces intractable divergence to word-level losses. |
On the Importance of the Kullback-Leibler Divergence Term in Variational Autoencoders for Text Generation (D19-56)
Copied to clipboard
| Challenge: | Variational Autoencoders suffer from learning uninformative latent representations due to issues such as approximated posterior collapse or entanglement of the latent space. |
| Approach: | They propose to impose an explicit constraint on the Kullback-Leibler divergence term inside the VAE objective function to understand the significance of the KL term in controlling the information transmitted through the VAe channel. |
| Outcome: | The proposed constraint avoids posterior collapse, but it also controls the information transmitted through the VAE channel. |
Dual-Space Knowledge Distillation for Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing large language models (LLMs) have strong generalization abilities due to their huge model capacities. |
| Approach: | They propose a dual-space knowledge distillation framework that unifies the output spaces of the two models for KD. |
| Outcome: | The proposed framework outperforms existing white-box KD frameworks on task-agnostic instruction-following benchmarks and can automatically align representations of two models with different vocabularies. |