Infini-Attention Boosts Long-Context Performance in Small Language Models

This paper explores the use of Infini-attention in small language models (SLMs) to improve their ability to handle long-context inputs. This is important because SLMs are more accessible and cost-effective than larger models, but often struggle with long sequences. The study provides empirical evidence that Infini-attention can significantly improve long-context retrieval accuracy in SLMs, even with limited parameters. The identification of the balance factor and the analysis of memory compression are valuable contributions to understanding the limitations and potential of this approach.