Researchers have developed smaller versions of large language models that can perform merchant information extraction in financial transactions with high accuracy. They fine-tuned various models, including LLaMA and Qwen, using LoRA (Large Model Adversarial Training) to reduce their size while preserving performance. The results show that compact models can achieve similar or even better accuracy than larger ones, but the choice of model depends on factors such as throughput and serving-stack behavior. --- Why it matters: This matters because it shows that large language models don't always need to be massive to perform well in specific tasks. Engineers can now explore smaller, more efficient models that balance performance with resource constraints, which is crucial for real-world applications like financial transactions. Source: https://arxiv.org/abs/2606.08051