Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Robustness of Agentic Function Calling

Ella Rabinovich, Ateret Anaby-Tavor

TL;DR

This work investigates the robustness of agentic function calling in large language models, addressing resilience to meaning-preserving input perturbations and toolkit expansion with related tools. It introduces a BFCL-based benchmark combining two perturbation axes—paraphrased user queries and expanded toolkits—to evaluate FC construction via AST-based correctness across multiple state-of-the-art models. The results reveal substantial fragility: many errors stem from parameter-value mismatches under exact-match evaluation, and toolkit expansion broadly degrades performance across models, highlighting gaps in current evaluation methods and real-world deployment challenges. The study emphasizes the need for semantic-aware evaluation and more diverse toolkits in agentic FC, outlining future work to broaden datasets and include non-relevant tools for more realistic robustness assessment.

Abstract

Large Language Models (LLMs) are increasingly acting as autonomous agents, with function calling (FC) capabilities enabling them to invoke specific tools for tasks. While prior research has primarily focused on improving FC accuracy, little attention has been given to the robustness of these agents to perturbations in their input. We introduce a benchmark assessing FC robustness in two key areas: resilience to naturalistic query variations, and stability in function calling when the toolkit expands with semantically related tools. Evaluating best-performing FC models on a carefully expanded subset of the Berkeley function calling leaderboard (BFCL), we identify critical weaknesses in existing evaluation methodologies, and highlight areas for improvement in real-world agentic deployments.

On the Robustness of Agentic Function Calling

TL;DR

This work investigates the robustness of agentic function calling in large language models, addressing resilience to meaning-preserving input perturbations and toolkit expansion with related tools. It introduces a BFCL-based benchmark combining two perturbation axes—paraphrased user queries and expanded toolkits—to evaluate FC construction via AST-based correctness across multiple state-of-the-art models. The results reveal substantial fragility: many errors stem from parameter-value mismatches under exact-match evaluation, and toolkit expansion broadly degrades performance across models, highlighting gaps in current evaluation methods and real-world deployment challenges. The study emphasizes the need for semantic-aware evaluation and more diverse toolkits in agentic FC, outlining future work to broaden datasets and include non-relevant tools for more realistic robustness assessment.

Abstract

Large Language Models (LLMs) are increasingly acting as autonomous agents, with function calling (FC) capabilities enabling them to invoke specific tools for tasks. While prior research has primarily focused on improving FC accuracy, little attention has been given to the robustness of these agents to perturbations in their input. We introduce a benchmark assessing FC robustness in two key areas: resilience to naturalistic query variations, and stability in function calling when the toolkit expands with semantically related tools. Evaluating best-performing FC models on a carefully expanded subset of the Berkeley function calling leaderboard (BFCL), we identify critical weaknesses in existing evaluation methodologies, and highlight areas for improvement in real-world agentic deployments.

Paper Structure

This paper contains 24 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Agentic Function Calling
  3. Robustness of Large Language Models
  4. Contribution
  5. Dataset Generation
  6. User Query Perturbations
  7. Expanding Agent's Toolkit
  8. (1)
  9. (2)
  10. (3)
  11. Agentic FC Robustness Evaluation
  12. Experimental Setup
  13. Models
  14. Evaluation Approach
  15. Experimental Results
  16. ...and 9 more sections

Figures (2)

  • Figure 1: Compact function definition example (top), and agent's output, triggering the function call with assigned parameter values (bottom), per user request "What is the record for the highest number of points scored by a single player in an NBA game?".
  • Figure 2: A toolkit expansion steps: (1) request variants are generated using the LLama3.1-70B model dubey2024llama, (2) function json definitions for executing these requests are generated using the Code-Llama-13B model roziere2023code, and a filtering step (3) is applied to filter out tools semantically identical to any of the original functions. The process is completed when the expanded toolkit is created for testing the original query.