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

TL;DR

This work investigates whether context aggregation patterns are shared across pretrained language models by introducing Order-Level Attention (OLA), derived from an order-wise decomposition of Attention Rollout. The authors show that same-order OLAs are highly similar across diverse LMs, a phenomenon termed OLAS, and demonstrate that OLAs encode syntactic knowledge, enabling a training-free cross-LM adapter transfer called TOA. TOA uses OLA as a unified syntactic feature and trains an adapter on a source LM that can be directly transferred to unseen LMs, yielding consistent improvements on relation extraction, named entity recognition, dependency parsing, and part-of-speech tagging, with stronger gains for smaller models and syntax-focused tasks. The findings advance understanding of shared LM mechanisms and offer a practical, low-cost pathway for cross-model knowledge transfer through cross-LM adapters. Code is released at the authors’ repository, enabling reproducibility and further exploration of OLAS and TOA.

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 https://github.com/jinglin-liang/OLAS.

Figures (7)

• Figure 1: Information flow path decomposition. (a) Simplified LM showing only Multi-Head Attention (MHA) modules (others omitted); (b) Path via residual connections across all layers; (c) Paths through MHA in one layer, residual connections elsewhere; (d) Path through MHA in all layer.
• Figure 2: Structure of each layer in typical LMs. (a) Llama and Qwen. (b) Bert, Roberta, and Electra. (c) Gemma.
• Figure 3: Visualization results of the Attention Rollout and first- and second-order OLA obtained by inputting two texts into Qwen2-1.5b and Llama3.2-3b. (a) Attention Rollout. (b) First-order OLA. (c) First-order OLA with row-wise maximum values set to zero. (d) Second-order OLA. (e) Second-order OLA with row-wise maximum values set to zero.
• Figure 4: Overview of TOA. In the training phase, the source LM is frozen and an adapter is trained for the downstream task using OLA as input. In the testing phase, the adapter is directly transferred to the target LM.
• Figure 5: Visualization results of Attention Rollout obtained from two texts input into Bert-base-cased, Roberta-base, Electra-base, Qwen2-1.5b, Gemma2-2b, and Llama3.2-3b. (a) Attention Rollout for Text 1. (b) Attention Rollout for Text 1 with outlier values set to zero. (c) Attention Rollout for Text 2. (d) Attention Rollout for Text 2 with outlier values set to zero.