On Evaluating the Integration of Reasoning and Action in LLM Agents with Database Question Answering
This work addresses the problem of assessing LLM agents' integration of reasoning and action for researchers, but it is incremental as it focuses on evaluation methods rather than novel agent capabilities.
The study tackled the challenge of evaluating LLMs in complex database question answering by introducing a new dataset requiring strategic SQL query generation and narrative synthesis, finding that even GPT-4 struggles with planning and multi-query generation bottlenecks.
This study introduces a new long-form database question answering dataset designed to evaluate how Large Language Models (LLMs) interact with a SQL interpreter. The task necessitates LLMs to strategically generate multiple SQL queries to retrieve sufficient data from a database, to reason with the acquired context, and to synthesize them into a comprehensive analytical narrative. Our findings highlight that this task poses great challenges even for the state-of-the-art GPT-4 model. We propose and evaluate two interaction strategies, and provide a fine-grained analysis of the individual stages within the interaction. A key discovery is the identification of two primary bottlenecks hindering effective interaction: the capacity for planning and the ability to generate multiple SQL queries. To address the challenge of accurately assessing answer quality, we introduce a multi-agent evaluation framework that simulates the academic peer-review process, enhancing the precision and reliability of our evaluations. This framework allows for a more nuanced understanding of the strengths and limitations of current LLMs in complex retrieval and reasoning tasks.