We present Uncertainty-Gated Lexical Decoding (UGLD), a decoding-time framework for fine-grained lexical control in Large Language Models (LLMs) that explicitly addresses the trade-off between controllability and fluency. UGLD adaptively scales intervention through an entropy-based gating mechanism derived from the model’s predictive distribution, activating control when uncertainty is high and limiting interference when predictions are confident. The method supports both promotion toward and against predefined vocabularies. We evaluate UGLD in Italian on two open-weight LLMs (ANITA 8B and Qwen 3 4B) across paraphrasing and free-text generation settings, considering Simple Vocabulary Conditioning and Jargon Reduction scenarios. Automatic evaluation shows consistent improvements in lexical coverage over standard decoding strategies, while human evaluation confirms that fluency is preserved under controlled intervention.

This paper presents a study on readability-controlled Sentence Simplification for Italian, addressing the scarcity of annotated resources for low-resource languages. We introduce IMPaCTS (Italian Multilevel Parallel Corpus for Text Simplification), the first fully automatically created corpus of 1,444,160 original–simple sentence pairs automatically annotated with readability levels and linguistic features. It was generated using an Italian LLM prompted in zero-shot to produce multiple simplifications per input sentence. Increasing portions of the resource are used to fine-tune mono- and multilingual open-weight LLMs, conditioning them to generate simplifications at a target readability level. Results from automatic and human evaluations show that fine-tuning on IMPaCTS improves performance both in terms of task completion and adherence to the targeted readability levels compared to few-shot baselines.

Recent advancements in Generative AI and Large Language Models (LLMs) have enabled the creation of highly realistic synthetic content, raising concerns about the potential for malicious use, such as misinformation and manipulation. Moreover, detecting Machine-Generated Text (MGT) remains challenging due to the lack of robust benchmarks that assess generalization to real-world scenarios. In this work, we evaluate the resilience of state-of-the-art MGT detectors (e.g., Mage, Radar, LLM-DetectAIve) to linguistically informed adversarial attacks. We develop a pipeline that fine-tunes language models using Direct Preference Optimization (DPO) to shift the MGT style toward human-written text (HWT), obtaining generations more challenging to detect by current models. Additionally, we analyze the linguistic shifts induced by the alignment and how detectors rely on “linguistic shortcuts” to detect texts. Our results show that detectors can be easily fooled with relatively few examples, resulting in a significant drop in detecting performances. This highlights the importance of improving detection methods and making them robust to unseen in-domain texts. We release code, models, and data to support future research on more robust MGT detection benchmarks.