KL-based self-distillation for large language models

• URL: http://arxiv.org/abs/2508.15807v1
• Date: Thu, 14 Aug 2025 15:45:50 GMT
• Title: KL-based self-distillation for large language models
• Authors: Max Rehman Linder,
• Abstract summary: We introduce a mathematically grounded method for knowledge distillation via KL divergence.<n>This allows the student model to inherit distributional knowledge from the teacher despite differing vocabularies.<n>We compare our KL-based distillation approach to conventional cross-entropy training, evaluating both methods for initializing new token embeddings.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large pre-trained language models often struggle to incorporate new domain-specific terminology when fine-tuned on small, specialized corpora. In this work, we address the challenge of vocabulary expansion in frozen LLMs by introducing a mathematically grounded method for knowledge distillation via KL divergence, even when the original and extended models use different tokenizations. This allows the student model to inherit distributional knowledge from the teacher despite differing vocabularies. We compare our KL-based distillation approach to conventional cross-entropy training, evaluating both methods across multiple strategies for initializing new token embeddings. After embedding initialization, models are further fine-tuned to integrate the new vocabulary. Each trained model is benchmarked on approximately 2000 code-generation tasks, where our approach achieves the best performance across the board. Finally, through mechanistic interpretability, we analyze how models learn representations for the new tokens, providing an explanation for the observed gains and offering insight into the structure of embedding space during vocabulary expansion.

Related papers

• Language Guided Concept Bottleneck Models for Interpretable Continual Learning [62.09201360376577]
  Continual learning aims to enable learning systems to acquire new knowledge constantly without forgetting previously learned information.<n>Most existing CL methods focus primarily on preserving learned knowledge to improve model performance.<n>We introduce a novel framework that integrates language-guided Concept Bottleneck Models to address both challenges.
  arXiv  Detail & Related papers  (2025-03-30T02:41:55Z)
• Scaling LLM Pre-training with Vocabulary Curriculum [0.0]
  We introduce vocabulary curriculum learning, an approach that improves pretraining efficiency with log-linear scaling gains relative to vocabulary size.<n>Our method alternates between entropy-guided vocabulary expansion and model optimization, enabling models to learn transferable representations across diverse tokenization granularities.<n>Experiments on small-scale GPT models demonstrate improved scaling efficiency, reinforcing the effectiveness of dynamic tokenization.
  arXiv  Detail & Related papers  (2025-02-25T07:18:29Z)
• The more polypersonal the better -- a short look on space geometry of fine-tuned layers [0.0]
  We analyze the changes in the internal representation of the BERT model when it is trained with additional grammatical modules.<n>We find that adding a single grammatical layer causes the model to separate the new and old grammatical systems within itself.
  arXiv  Detail & Related papers  (2025-01-09T18:50:47Z)
• Large Concept Models: Language Modeling in a Sentence Representation Space [62.73366944266477]
  We present an attempt at an architecture which operates on an explicit higher-level semantic representation, which we name a concept.<n> Concepts are language- and modality-agnostic and represent a higher level idea or action in a flow.<n>We show that our model exhibits impressive zero-shot generalization performance to many languages.
  arXiv  Detail & Related papers  (2024-12-11T23:36:20Z)
• A Sentence Speaks a Thousand Images: Domain Generalization through Distilling CLIP with Language Guidance [41.793995960478355]
  We propose a novel approach for domain generalization that leverages recent advances in large vision-language models. The key technical contribution is a new type of regularization that requires the student's learned image representations to be close to the teacher's learned text representations. We evaluate our proposed method, dubbed RISE, on various benchmark datasets and show that it outperforms several state-of-the-art domain generalization methods.
  arXiv  Detail & Related papers  (2023-09-21T23:06:19Z)
• On Conditional and Compositional Language Model Differentiable Prompting [75.76546041094436]
  Prompts have been shown to be an effective method to adapt a frozen Pretrained Language Model (PLM) to perform well on downstream tasks. We propose a new model, Prompt Production System (PRopS), which learns to transform task instructions or input metadata, into continuous prompts.
  arXiv  Detail & Related papers  (2023-07-04T02:47:42Z)
• Autoregressive Structured Prediction with Language Models [73.11519625765301]
  We describe an approach to model structures as sequences of actions in an autoregressive manner with PLMs. Our approach achieves the new state-of-the-art on all the structured prediction tasks we looked at.
  arXiv  Detail & Related papers  (2022-10-26T13:27:26Z)
• Language Model Pre-Training with Sparse Latent Typing [66.75786739499604]
  We propose a new pre-training objective, Sparse Latent Typing, which enables the model to sparsely extract sentence-level keywords with diverse latent types. Experimental results show that our model is able to learn interpretable latent type categories in a self-supervised manner without using any external knowledge.
  arXiv  Detail & Related papers  (2022-10-23T00:37:08Z)
• Interpreting Language Models Through Knowledge Graph Extraction [42.97929497661778]
  We compare BERT-based language models through snapshots of acquired knowledge at sequential stages of the training process. We present a methodology to unveil a knowledge acquisition timeline by generating knowledge graph extracts from cloze "fill-in-the-blank" statements. We extend this analysis to a comparison of pretrained variations of BERT models (DistilBERT, BERT-base, RoBERTa)
  arXiv  Detail & Related papers  (2021-11-16T15:18:01Z)
• SLM: Learning a Discourse Language Representation with Sentence Unshuffling [53.42814722621715]
  We introduce Sentence-level Language Modeling, a new pre-training objective for learning a discourse language representation. We show that this feature of our model improves the performance of the original BERT by large margins.
  arXiv  Detail & Related papers  (2020-10-30T13:33:41Z)

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.