Refining User Instructions for Better LLM Assistance

Anonymous

TL;DR: use a refiner to deal with flawed user instruction in llm assistant application

Abstract:

A typical application scenario for generative LLMs is directly interacting with end-users in conversation. However, the distribution of actual user instructions can differ from those in the publicly available datasets, which could negatively influence the user experience.
In this paper, we propose a new method to overcome the instruction's difference via regenerating the instruction. We address a specific case of how user instruction can differ: more flaws can exist in their daily expressions. We leverage instruction-tuned LLMs to refine the flawed instruction so they better align with the training distribution. We explored the effectiveness of directly asking the model to refine the instruction and further finetuned a specialized refiner model to enhance the overall performance.
Our experiments demonstrate the effectiveness of the proposed method on the open-source model, especially when using a finetuned model as the refiner. The enhancement is achieved without requiring retraining or parameter increasing on the assistant model, highlighting its practicality and potential to bridge the gap between open-source and proprietary LLM assistants.

Paper Type: short

Research Area: NLP Applications

Contribution Types: NLP engineering experiment

Languages Studied: English