Enhancing LLM Instruction Following: An Evaluation-Driven Multi-Agentic Workflow for Prompt Instructions Optimization

TL;DR

The paper tackles reliable instruction following by decoupling the primary task from its constraints and optimizing the constraints via quantitative feedback. It presents a multi-agent, evaluation-driven workflow that iteratively rewrites explicit constraints using constraint-specific scores and a planner–executor architecture. An explicit constraints extraction step extends InfoBench to generate modular constraints, and the framework operates through a four-stage cycle: Content Generation, Evaluation, Action Planning, and Constraint Editing. Empirical results on prompts for Llama 3.1 8B and Mixtral-8x 7B show substantial improvements in compliance, with a baseline around 82% rising to about 91.5–91.6%, demonstrating practical potential for automated constraint enforcement in real-world systems.

Abstract

Large Language Models (LLMs) often generate substantively relevant content but fail to adhere to formal constraints, leading to outputs that are conceptually correct but procedurally flawed. Traditional prompt refinement approaches focus on rephrasing the description of the primary task an LLM has to perform, neglecting the granular constraints that function as acceptance criteria for its response. We propose a novel multi-agentic workflow that decouples optimization of the primary task description from its constraints, using quantitative scores as feedback to iteratively rewrite and improve them. Our evaluation demonstrates this method produces revised prompts that yield significantly higher compliance scores from models like Llama 3.1 8B and Mixtral-8x 7B.

Figures (1)

• Figure 1: Proposed workflow for instruction refinement.