Dual-Space Knowledge Distillation with Key-Query Matching for Large Language Models with Vocabulary Mismatch
This work addresses efficiency deployment issues for large language models by improving cross-tokenizer knowledge distillation, but it is incremental as it builds on existing methods with specific enhancements.
The paper tackled the problem of knowledge distillation between large language models with different tokenizers by analyzing a state-of-the-art method and proposing a novel generative adversarial approach, resulting in modest ROUGE-L gains of +0.37 on average for text generation quality, especially on out-of-distribution data.
Large language models (LLMs) achieve state-of-the-art (SOTA) performance across language tasks, but are costly to deploy due to their size and resource demands. Knowledge Distillation (KD) addresses this by training smaller Student models to mimic larger Teacher models, improving efficiency without significant performance loss. Dual-Space Knowledge Distillation with Cross-Model Attention (DSKD-CMA) has emerged as a SOTA method for KD between LLMs with distinct tokenizers, yet its internal workings remain largely opaque. In this work, we systematically analyse the attention mechanism of DSKD-CMA through manual token alignment probing and heatmap visualisations, revealing both strengths and limitations. Building on this, we introduce a novel method, DSKD-CMA-GA, based on Generative Adversarial (GA) learning, to address the mismatched distributions between the keys and queries computed from distinct models. Experiments show modest but consistent ROUGE-L gains in text generation quality, particularly on out-of-distribution data (+0.37 on average), narrowing the gap between cross- and same-tokenizer KD.