Reproduce “Automated test generation to evaluate tool-augmented LLMs as conversational AI agents” on OpenResearch

About this paper

Samuel Arcadinho, David Aparicio, Mariana Almeida

This research from Zendesk introduces an automated pipeline for generating evaluation datasets for tool-augmented conversational AI agents, particularly for customer support. The methodology, which employs intermediate graph structures and noise injection, successfully produced the ALMITA dataset and revealed that while current LLMs perform well on isolated actions, they struggle to maintain accuracy over entire multi-turn, procedure-driven conversations, with conversation correctness ranging from 1.6% to 15.6% across tested models.

Key insights

• Intermediate graph structures (flowgraphs and conversation graphs) are critical for generating valid, complete, and low-hallucination conversational test data for LLM agents.
• Current tool-augmented LLMs exhibit very low 'conversation correctness' (1.6% to 15.6%), indicating they frequently fail to maintain accuracy over multi-turn, procedure-driven dialogues, despite reasonable single-turn performance.
• An automated pipeline can generate evaluation datasets that accurately rank LLM agent performance, even without extensive manual curation, which is crucial for scalable and continuous evaluation.

Summary by alphaXiv.