Scholarlink

Large Language Models (LLMs) have demonstrated remarkable performance in natural language processing tasks like writing. However, LLMs are consuming amount of computing resources to architect the state-of-the-art performance, which will limit the scope of application of LLMs, as real-world devices are mostly incapable of handling the load imposed by LLMs training and inferencing. Nowadays, Tiny-LLMs give a chance that LLMs no longer have as huge parameters as the last LLMs did, which are easy to use and can be tailored in specific task. In this work, we consider the most favorite LLMs: Llama-2-42M and Llama-3.2-1B, which be employed with prompt engineering, sampling technique, top- \({\text{k}}\) and top- \({\text{p}}\) algorithms in writing. And we propose a scheme for boosting the performance of Tiny-LLM that can generate long stories stably. The experimental results are evaluated in perplexity.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Can Tiny Tell Stories Well? Exploiting Tiny-LLMs Performance in Writing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hui Ding,
  • Guokun Cui,
  • Minghe Xu

摘要

Large Language Models (LLMs) have demonstrated remarkable performance in natural language processing tasks like writing. However, LLMs are consuming amount of computing resources to architect the state-of-the-art performance, which will limit the scope of application of LLMs, as real-world devices are mostly incapable of handling the load imposed by LLMs training and inferencing. Nowadays, Tiny-LLMs give a chance that LLMs no longer have as huge parameters as the last LLMs did, which are easy to use and can be tailored in specific task. In this work, we consider the most favorite LLMs: Llama-2-42M and Llama-3.2-1B, which be employed with prompt engineering, sampling technique, top- \({\text{k}}\) and top- \({\text{p}}\) algorithms in writing. And we propose a scheme for boosting the performance of Tiny-LLM that can generate long stories stably. The experimental results are evaluated in perplexity.