English

Convolution for Large Language Models

Computation and Language 2026-07-20 v1

Abstract

Large language models (LLMs) largely rely on Transformers, where self-attention provides global token interaction but does not explicitly encode the locality of natural language. We study whether lightweight depthwise convolutions can supply this local inductive bias without materially increasing model size. Our macro-level ablation compares convolution at 17 locations in a Qwen3 Transformer block and finds the best results when convolution is applied to the projected queries, keys, and values before attention. A subsequent micro-level study favors a residual depthwise convolution with kernel size k=3k=3, without additional normalization or activation. Across Qwen3 models and several pre-training data budgets, this design improves the average accuracy on seven downstream benchmarks while adding less than 0.01%0.01\% parameters. A representation-level case study further suggests that the convolution makes repeated token IDs more sensitive to their immediate context. These results support depthwise convolution as a lightweight complement to self-attention for modeling short-range token interactions.

Cite

@article{arxiv.2607.18413,
  title  = {Convolution for Large Language Models},
  author = {Yuchuan Tian and Yingte Shu and Wei He and Shuo Zhang and Tianchen Zhao and Chao Xu and Xinghao Chen and Yunhe Wang and Hanting Chen and Yu Wang},
  journal= {arXiv preprint arXiv:2607.18413},
  year   = {2026}
}

Comments

12 pages, 5 figures