SRA: Span Representation Alignment for Large Language Model Distillation

TL;DR

This paper introduces SRA, a novel span representation alignment framework for large language model distillation that improves cross-architecture knowledge transfer by focusing on robust, tokenizer-agnostic spans modeled as multi-particle systems.

Contribution

SRA redefines token alignment by using semantic-rich spans as the fundamental unit, employing a physical analogy and geometric regularization to enhance distillation effectiveness.

Findings

SRA outperforms state-of-the-art CTKD methods in cross-architecture distillation.

Modeling spans as multi-particle systems improves semantic robustness.

Attention-weighted span centers of mass enhance knowledge transfer.

Abstract

Cross-Tokenizer Knowledge Distillation (CTKD) enables knowledge transfer between a large language model and a smaller student, even when they employ different tokenizers. While existing approaches mainly focus on token-level alignment strategies, which are often brittle and sensitive to discrepancies between tokenizers, we argue that the method of aggregating tokens into more robust representations before distillation is of equal importance. In this paper, we introduce \textbf{SRA} (\textbf{S}pan \textbf{R}epresentation \textbf{A}lignment for Large Language Model Distillation), a novel framework that reframes CTKD through the physical lens of Multi-Particle Dynamical Systems. SRA shifts the fundamental unit of alignment from tokens to robust, tokenizer-agnostic spans. We model each span as a cluster of particles and represent its state by its Center of Mass (CoM) - an attention-weighted average that captures rich semantic information. We leverage the concept of span centers of mass with attention-derived weighting to prioritize the most salient spans. In addition, we employ a geometric regularizer to preserve the structural integrity of the representation space and introduce aligned span logit distillation to enhance knowledge transfer across models. In challenging cross-architecture distillation experiments, SRA consistently and significantly outperforms state-of-the-art CTKD baselines, validating our physically-grounded approach.