Researchers from Carson Rodrigues and Oysturn Vas claim that large language models (LLMs) can replace fine-tuned natural language understanding (NLU) classifiers for intent detection in production conversational systems. However, their study suggests that the answer is not a simple yes or no. Instead, it depends on the specific characteristics of the intent space. In some cases, LLMs outperform fine-tuned NLU classifiers, but in others, they are tied or even worse. The researchers identify three production-relevant regimes where LLMs have an advantage: out-of-scope detection, robustness to realistic ASR noise, and dynamic per-deployment schemas. They propose a decision framework for practitioners to help them choose between LLMs and fine-tuned NLU classifiers based on their specific needs. --- Why it matters: This study matters because it provides practical guidance for engineers building conversational systems, helping them decide when to use LLMs versus fine-tuned NLU classifiers. The findings can inform the development of more efficient and effective intent detection systems. Source: https://arxiv.org/abs/2608.20371