Generative Prompt Internalization

TL;DR

Generative Prompt Internalization (GenPI) is a lightweight training method that internalizes complex prompts into language models, reducing computational overhead and enabling efficient inference without explicit prompts.

Contribution

The paper introduces GenPI, a novel joint training approach that internalizes prompts and their reasoning into models, along with a data synthesis technique for training without environment interactions.

Findings

Effectively internalizes complex prompts across various scenarios.

Reduces inference costs by eliminating explicit prompts.

Maintains high performance without prompt input.

Abstract

Prompts used in recent large language model based applications are often fixed and lengthy, leading to significant computational overhead. To address this challenge, we propose Generative Prompt Internalization (GenPI), a lightweight method that employs a joint training approach. GenPI not only replicates the behavior of models with prompt inputs but also generates the content of the prompt along with reasons for why the model's behavior should change accordingly. We demonstrate that our approach effectively internalizes complex prompts across various agent-based application scenarios. For effective training without interactions with the dedicated environments, we introduce a data synthesis technique that autonomously collects conversational datasets by swapping the roles of the agent and environment. This method is especially useful in scenarios where only a predefined prompt is available without a corresponding training dataset. By internalizing complex prompts, Generative Prompt Internalization enables high performance and efficient inference without the need for explicit prompts.