GOAT: A Training Framework for Goal-Oriented Agent with Tools

TL;DR

GOAT introduces an annotation-free training framework that builds synthetic goal-oriented API execution data directly from API documents by constructing and refining an API dependency graph. It enables domain-specific fine-tuning of open-source LLM agents to plan and execute interdependent API calls, paired with a retriever, resulting in state-of-the-art performance on multiple goal-oriented benchmarks. The authors also present GOATBench, a large-scale evaluation suite spanning numerous tools and APIs to test generalization across seen and unseen setups. Empirical results demonstrate strong gains for GOAT-trained models across diverse backbones, underscoring the practical potential of automatic data generation for complex tool use. Overall, GOAT offers a scalable path to robust, open-source agents capable of sophisticated reasoning and multi-step tool orchestration.

Abstract

Large language models (LLMs) have recently been extended beyond traditional text generation to serve as interactive agents capable of using external tools based on user intent. However, current LLM agents still show limited ability to handle goal-oriented queries, which require decomposing a high-level objective into multiple interdependent API calls with correct planning and execution. Current approaches mainly rely on zero-shot evaluation due to the absence of training data. While proprietary closed-source models such as GPT-4 demonstrate strong reasoning abilities, smaller open-source models struggle to perform complex tool use effectively. Thus, we propose a novel training framework GOAT, which enables fine-tuning of LLM agents in a human annotation-free setting. GOAT automatically constructs synthetic datasets of goal-oriented API execution tasks directly from given API documents, equipping models with the ability to reason over interdependent calls and generate coherent responses. Through extensive experiments, we show that GOAT-trained agents achieve state-of-the-art performance across multiple existing goal-oriented benchmarks. In addition, we introduce GOATBench, a new goal-oriented API execution benchmark, and demonstrate that agents trained with GOAT also excel in this setting. These results highlight GOAT as a practical path toward building robust open-source LLM agents capable of complex reasoning and tool use.

Figures (21)

• Figure 1: Goal-oriented API execution task. To solve a goal-oriented user query, the LLM agent performs step-by-step task planning, executes a sequence of interdependent API calls, and generates a natural language response. The figure illustrates the workflow where the user query is decomposed into subtasks, mapped to API calls, and each function call is executed by filling API arguments based on the outputs of previous responses, ultimately yielding a coherent answer.
• Figure 2: The overview of API dependency graph construction process. Given the API documents, each document is first parsed to extract function descriptions, which are then used to initialize a raw dependency graph in (a). This graph is progressively refined through three filtering steps (c)-(e), resulting in the final API dependency graph that captures reliable relations among APIs. The graphs shown under (b)-(e) illustrate how the API dependency graph evolves as it is progressively refined through each filtering step.
• Figure 3: Overview of goal-oriented API execution data construction. The process involves (a) sampling connected API sequences, (b) generating API calls, outputs, and sub-queries, and (c) composing user queries and final responses.
• Figure 4: Example of API document parsing result.
• Figure 5: Example of Constructed API Dependency Graph from APIBank APIs.