Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Huang, Ruizhe, Zhang, Kexuan, Fang, Yihao, Yu, Baifeng
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2512.23862
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914225032527872
author Huang, Ruizhe
Zhang, Kexuan
Fang, Yihao
Yu, Baifeng
author_facet Huang, Ruizhe
Zhang, Kexuan
Fang, Yihao
Yu, Baifeng
contents This study investigates small-scale pretraining for Small Language Models (SLMs) to enable efficient use of limited data and compute, improve accessibility in low-resource settings and reduce costs. To enhance long-context extrapolation in compact models, we focus on Infini-attention, which builds a compressed memory from past segments while preserving local attention. In our work, we conduct an empirical study using 300M-parameter LLaMA models pretrained with Infini-attention. The model demonstrates training stability and outperforms the baseline in long-context retrieval. We identify the balance factor as a key part of the model performance, and we found that retrieval accuracy drops with repeated memory compressions over long sequences. Even so, Infini-attention still effectively compensates for the SLM's limited parameters. Particularly, despite performance degradation at a 16,384-token context, the Infini-attention model achieves up to 31% higher accuracy than the baseline. Our findings suggest that achieving robust long-context capability in SLMs benefits from architectural memory like Infini-attention.
format Preprint
id arxiv_https___arxiv_org_abs_2512_23862
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Probing the Limits of Compressive Memory: A Study of Infini-Attention in Small-Scale Pretraining
Huang, Ruizhe
Zhang, Kexuan
Fang, Yihao
Yu, Baifeng
Machine Learning
Artificial Intelligence
Computation and Language
This study investigates small-scale pretraining for Small Language Models (SLMs) to enable efficient use of limited data and compute, improve accessibility in low-resource settings and reduce costs. To enhance long-context extrapolation in compact models, we focus on Infini-attention, which builds a compressed memory from past segments while preserving local attention. In our work, we conduct an empirical study using 300M-parameter LLaMA models pretrained with Infini-attention. The model demonstrates training stability and outperforms the baseline in long-context retrieval. We identify the balance factor as a key part of the model performance, and we found that retrieval accuracy drops with repeated memory compressions over long sequences. Even so, Infini-attention still effectively compensates for the SLM's limited parameters. Particularly, despite performance degradation at a 16,384-token context, the Infini-attention model achieves up to 31% higher accuracy than the baseline. Our findings suggest that achieving robust long-context capability in SLMs benefits from architectural memory like Infini-attention.
title Probing the Limits of Compressive Memory: A Study of Infini-Attention in Small-Scale Pretraining
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2512.23862