Don't Ignore the Tail: Decoupling top-K Probabilities for Efficient Language Model Distillation

• URL: http://arxiv.org/abs/2602.20816v1
• Date: Tue, 24 Feb 2026 11:54:06 GMT
• Title: Don't Ignore the Tail: Decoupling top-K Probabilities for Efficient Language Model Distillation
• Authors: Sayantan Dasgupta, Trevor Cohn, Timothy Baldwin,
• Abstract summary: We propose a new tail-aware divergence that decouples the contribution of the teacher model's top-K predicted probabilities from that of lower-probability predictions.<n> Experimental results demonstrate that our modified distillation method yields competitive performance in both pre-training and supervised distillation of decoder models.
• Score: 50.19746127327559
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: The core learning signal used in language model distillation is the standard Kullback-Leibler (KL) divergence between the student and teacher distributions. Traditional KL divergence tends to be dominated by the next tokens with the highest probabilities, i.e., the teacher's modes, thereby diminishing the influence of less probable yet potentially informative components of the output distribution. We propose a new tail-aware divergence that decouples the contribution of the teacher model's top-K predicted probabilities from that of lower-probability predictions, while maintaining the same computational profile as the KL Divergence. Our decoupled approach reduces the impact of the teacher modes and, consequently, increases the contribution of the tail of the distribution. Experimental results demonstrate that our modified distillation method yields competitive performance in both pre-training and supervised distillation of decoder models across various datasets. Furthermore, the distillation process is efficient and can be performed with a modest academic budget for large datasets, eliminating the need for industry-scale computing.

Related papers

• Knowledge Distillation of Uncertainty using Deep Latent Factor Model [10.148306002388196]
  We introduce a new method of distribution distillation called Gaussian distillation.<n>It estimates the distribution of a teacher ensemble through a special Gaussian process called the deep latent factor model (DLF)<n>By using multiple benchmark datasets, we demonstrate that the proposed Gaussian distillation outperforms existing baselines.
  arXiv  Detail & Related papers  (2025-10-22T06:46:59Z)
• ToDi: Token-wise Distillation via Fine-Grained Divergence Control [9.958797874295355]
  Token-wise Distillation (ToDi) is a novel method that adaptively combines Forward KL and Reverse KL per token using a sigmoid-based weighting function.<n>ToDi consistently outperforms recent distillation baselines using uniform or less granular strategies.
  arXiv  Detail & Related papers  (2025-05-22T06:51:16Z)
• A Generative Framework for Causal Estimation via Importance-Weighted Diffusion Distillation [55.53426007439564]
  Estimating individualized treatment effects from observational data is a central challenge in causal inference.<n>In inverse probability weighting (IPW) is a well-established solution to this problem, but its integration into modern deep learning frameworks remains limited.<n>We propose Importance-Weighted Diffusion Distillation (IWDD), a novel generative framework that combines the pretraining of diffusion models with importance-weighted score distillation.
  arXiv  Detail & Related papers  (2025-05-16T17:00:52Z)
• Enhancing Knowledge Distillation of Large Language Models through Efficient Multi-Modal Distribution Alignment [10.104085497265004]
  We propose Ranking Loss based Knowledge Distillation (RLKD), which encourages consistency of peak predictions between the teacher and student models.<n>Our method enables the student model to better learn the multi-modal distributions of the teacher model, leading to a significant performance improvement in various downstream tasks.
  arXiv  Detail & Related papers  (2024-09-19T08:06:42Z)
• 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)
• Knowledge Distillation Performs Partial Variance Reduction [93.6365393721122]
  Knowledge distillation is a popular approach for enhancing the performance of ''student'' models. The underlying mechanics behind knowledge distillation (KD) are still not fully understood. We show that KD can be interpreted as a novel type of variance reduction mechanism.
  arXiv  Detail & Related papers  (2023-05-27T21:25:55Z)
• Churn Reduction via Distillation [54.5952282395487]
  We show an equivalence between training with distillation using the base model as the teacher and training with an explicit constraint on the predictive churn. We then show that distillation performs strongly for low churn training against a number of recent baselines.
  arXiv  Detail & Related papers  (2021-06-04T18:03:31Z)
• Scaling Ensemble Distribution Distillation to Many Classes with Proxy Targets [12.461503242570643]
  emphEnsemble Distribution Distillation is an approach that allows a single model to efficiently capture both the predictive performance and uncertainty estimates of an ensemble. For classification, this is achieved by training a Dirichlet distribution over the ensemble members' output distributions via the maximum likelihood criterion. Although theoretically, this criterion exhibits poor convergence when applied to large-scale tasks where the number of classes is very high.
  arXiv  Detail & Related papers  (2021-05-14T17:50:14Z)
• 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.