Prefix-tuning is a lightweight alternative to fine-tuning for natural language generation tasks. It keeps language model parameters frozen and optimizes a small continuous task-specific vector called a prefix. The prefix is treated as "virtual tokens" that subsequent tokens can attend to, allowing the model to generate text based on the prefix without modifying the original model. This approach is more efficient in terms of storage and computation, as it only requires optimizing the prefix parameters, which are much fewer than the full model parameters. Prefix-tuning has been applied to GPT-2 for table-to-text generation and to BART for summarization. It achieves comparable performance to fine-tuning in full data settings and outperforms fine-tuning in low-data settings and extrapolation to unseen topics. Prefix-tuning is modular and space-efficient, allowing for easy adaptation to multiple tasks without modifying the original model. It also enables processing examples from multiple users or tasks in a single batch, avoiding data cross-contamination. The method is effective in both table-to-text and summarization tasks, with prefix-tuning showing better performance in low-data and extrapolation scenarios. The results demonstrate that prefix-tuning is a promising approach for adapting large language models to various generation tasks with minimal computational overhead.Prefix-tuning is a lightweight alternative to fine-tuning for natural language generation tasks. It keeps language model parameters frozen and optimizes a small continuous task-specific vector called a prefix. The prefix is treated as "virtual tokens" that subsequent tokens can attend to, allowing the model to generate text based on the prefix without modifying the original model. This approach is more efficient in terms of storage and computation, as it only requires optimizing the prefix parameters, which are much fewer than the full model parameters. Prefix-tuning has been applied to GPT-2 for table-to-text generation and to BART for summarization. It achieves comparable performance to fine-tuning in full data settings and outperforms fine-tuning in low-data settings and extrapolation to unseen topics. Prefix-tuning is modular and space-efficient, allowing for easy adaptation to multiple tasks without modifying the original model. It also enables processing examples from multiple users or tasks in a single batch, avoiding data cross-contamination. The method is effective in both table-to-text and summarization tasks, with prefix-tuning showing better performance in low-data and extrapolation scenarios. The results demonstrate that prefix-tuning is a promising approach for adapting large language models to various generation tasks with minimal computational overhead.