ELiRF-UPV at SimpleText 2026: Decoupling Action Prediction and Controlled Generation for Text Simplification

In this work, we describe the ELiRF-UPV participation in the CLEF 2026 SimpleText Track, Simplify Scientific Text. Our approach for text simplification decouples the process into two stages: planning and generation. First, a model predicts sentence-level simplification actions, which are then used to guide the final simplification. For the generation stage, we explore and compare two primary strategies. The first one is LLM-based, where we apply prompting techniques. For the second strategy, we fine-tune an encoder-decoder-based automatic simplification model. This design allows the models to learn new actions without suffering from catastrophic forgetting of their pre-trained knowledge. In addition, we apply Direct Preference Optimization (DPO) training to strictly align the model’s outputs with the predicted control instructions, mitigating unwarranted changes and ensuring accurate simplifications. A core principle of our work is its alignment with Green AI guidelines; by using small open-source models, we aim to democratize controlled text generation and reduce the high computational and environmental costs associated with massive proprietary models.