LLM-based text simplification with CEFR-level control in TSAR @ EMNLP 2025

HumanAI researchers from the HULAT group (UC3M) have taken part in the  4th edition of the TSAR Shared Task on Readability-Controlled Text Simplification, held in conjunction with the 30th EMNLP conference (Empirical Methods in Natural Language Processing) in Suzhou, China.

The team presented the paper “HULAT-UC3M at TSAR 2025 Shared Task: A Prompt-Based Approach using Lightweight Language Models for Readability-Controlled Text Simplification”, which explores how lightweight Large Language Models (LLMs) can be used to adapt English texts to different readability levels defined by the CEFR (A1, A2, B1).

The proposed system combines models such as LLaMA and Ettin Decoder with two alternative prompt engineering strategies (one more detailed, one more concise). The goal is to simplify original texts at B2 level or higher and produce versions that are easier to understand for different reader profiles, while preserving meaning and essential information.

For HumanAI, this participation is directly aligned with the project’s mission to better understand how to control reading level, and identifying where LLMs still fail to adapt content reliably is crucial for future interfaces that can automatically present information in a way that matches each user’s cognitive needs.

A more detailed description of the system and results can be found in the paper published in the TSAR 2025 proceedings.

• Jesus M. Sanchez-Gomez, Lourdes Moreno, Paloma Martínez, and Marco Antonio Sanchez-Escudero. 2025. HULAT-UC3M at TSAR 2025 Shared Task A Prompt-Based Approach using Lightweight Language Models for Readability-Controlled Text Simplification. In Proceedings of the Fourth Workshop on Text Simplification, Accessibility and Readability (TSAR 2025), pages 183–192, Suzhou, China. Association for Computational Linguistics.