LLM API Calling Fails in Multilingual Settings: Study Proposes Post-Training Fixes

A recent investigation published on arXiv (2608.11715) explores the dependability of Large Language Models (LLMs) in making API calls across multiple languages. The researchers highlight a frequent issue known as 'Argument Language Mismatch' (ALM), where a model correctly identifies the tool but produces argument values in a different language. While these outputs are semantically valid, they are functionally incorrect and not measured by traditional API-calling metrics. The study reassesses post-training techniques to alleviate ALM, revealing that supervised fine-tuning (SFT) establishes a robust baseline, greatly enhancing argument language uniformity and overall function call precision. When models are consistently selected, SFT's performance rivals or even surpasses that of more intricate reinforcement learning (RL) methods. The research also considers whether RL with structured, argument-aware rewards provides added advantages, observing that techniques like Group Relative Policy Optimization (GRPO) can enhance specific elements. These results imply that simpler post-training approaches might effectively tackle multilingual tool usage, questioning the need for complex RL frameworks.