Order-Level Attention Similarity Across Language Models: A Latent Commonality

Jinglin Liang, Jin Zhong, Shuangping Huang, Yunqing Hu, Huiyuan Zhang, Huifang Li, Lixin Fan, Hanlin Gu

Keywords: Exploration of language model commonalities, Language model adapter transfer

TL;DR: We discover commonalities across different language models and leverage them to achieve training-free cross-model adapter transfer.

Abstract:

In this paper, we explore an important yet previously neglected question: Do context aggregation patterns across Language Models (LMs) share commonalities?
While some works have investigated context aggregation or attention weights in LMs, they typically focus on individual models or attention heads, lacking a systematic analysis across multiple LMs to explore their commonalities.
In contrast, we focus on the commonalities among LMs, which can deepen our understanding of LMs and even facilitate cross-model knowledge transfer.
In this work, we introduce the Order-Level Attention (OLA) derived from the order-wise decomposition of Attention Rollout and reveal that the OLA at the same order across LMs exhibits significant similarities.
Furthermore, we discover an implicit mapping between OLA and syntactic knowledge.
Based on these two findings, we propose the Transferable OLA Adapter (TOA), a training-free cross-LM adapter transfer method.
Specifically, we treat the OLA as a unified syntactic feature representation and train an adapter that takes OLA as input.
Due to the similarities in OLA across LMs, the adapter generalizes to unseen LMs without requiring any parameter updates.
Extensive experiments demonstrate that TOA's cross-LM generalization effectively enhances the performance of unseen LMs.
Code is available at \url{https://github.com/jinglin-liang/OLAS}.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 11941