.jpg)
Automatic generation of BPMN models from natural language descriptions has advanced thanks to large language models (LLMs). However, traditional supervised fine-tuning limits quality to the patterns of the training set. This is where reinforcement learning (RL) comes into play, enabling optimization beyond that ceiling through carefully designed reward functions. A recent study explores how to design these rewards when quality is multidimensional —syntactic, pragmatic, and semantic— training models such as Llama 3.1 8B and Qwen 2.5 14B under 48 configurations. The results reveal that RL significantly improves pragmatic and syntactic quality, reduces variability by more than six times, and that equal weighting of rewards outperforms any specific approach. Furthermore, interactions with the model architecture are complex: an invalidity penalty is essential for one model but irrelevant for another, and initialization with supervised fine-tuning can be counterproductive. These findings are applicable to any structured generation task where quality is measured across multiple automated dimensions.
For companies looking to integrate these capabilities into their workflows, having custom applications that incorporate artificial intelligence is key. At Q2BSTUDIO we offer AI for businesses that allows training models with RL to automate business modeling processes, improving accuracy and reducing manual intervention. Additionally, our process automation solutions are complemented by AWS and Azure cloud services to scale the training and deployment of these models. Cybersecurity is also a priority when handling sensitive data during model generation, so we integrate penetration testing and information assurance. For result analysis, we use business intelligence services with Power BI, visualizing performance metrics of AI agents. All of this is part of an ecosystem of custom software that drives digital transformation.
