AvaTaR: Optimizing LLM Agents for Tool Usage via Contrastive Reasoning

TL;DR

A novel and automated framework that optimizes an LLM agent to effectively leverage provided tools, improving performance on a given task is introduced, and AvaTaR consistently outperforms state-of-the-art approaches.

Abstract

Large language model (LLM) agents have demonstrated impressive capabilities in utilizing external tools and knowledge to boost accuracy and reduce hallucinations. However, developing prompting techniques that enable LLM agents to effectively use these tools and knowledge remains a heuristic and labor-intensive task. Here, we introduce AvaTaR, a novel and automated framework that optimizes an LLM agent to effectively leverage provided tools, improving performance on a given task. During optimization, we design a comparator module to iteratively deliver insightful and comprehensive prompts to the LLM agent by contrastively reasoning between positive and negative examples sampled from training data. We demonstrate AvaTaR on four complex multimodal retrieval datasets featuring textual, visual, and relational information, and three general question-answering (QA) datasets. We find AvaTaR consistently outperforms state-of-the-art approaches across all seven tasks, exhibiting strong generalization ability when applied to novel cases and achieving an average relative improvement of 14% on the Hit@1 metric for the retrieval datasets and 13% for the QA datasets. Code and dataset are available at https://github.com/zou-group/avatar.

Figures (8)

• Figure 1: Overview of AvaTaR.AvaTaR consists of a actor LLM and a comparator LLM. (a) During optimization, the actor generates actions to answer queries by leveraging the provided tools. Then, the comparator contrasts a set of well-performing (positive) and poorly-performing (negative) queries, automatically generating holistic prompts to teach the actor more effective retrieval strategies and tool usage (cf. Section \ref{['sec:method']}). (b) At deployment, the actor with optimized prompts or actions can be effectively used to answer new queries.
• Figure 2: Comparison between AvaTaR and ReAct. (a) The ReAct agent exhibits incomplete task decomposition and employs suboptimal tool combinations, such as lengthy string matching, leading to poor task performance. (b) AvaTaR decomposes the task into multiple steps, such as type filtering and flexible token matching. Moreover, it implements robust tool usage and precise synthesis with learned parameters from the optimization phase to achieve excellent performance on new queries.
• Figure 3: Demonstration example during optimization. Best viewed in color. The task of the comparator is to automatically generate instructions based on sampled positive and negative queries. Then comparator provides holistic instructions that guide the actor to improve query decomposition, utilize better tools, and incorporate more comprehensive information.
• Figure 4: Optimization dynamics of AvaTaR agents on STaRK. The figures show validation performance (solid line) and its moving average (dashed line) during the optimization of AvaTaR.
• Figure 5: Performance (left) and AvaTaR's optimization dynamics (right) on Flickr30K-Entities.