Sébastien Bratières

Michele Miranda, Elena Sofia Ruzzetti, Andrea Santilli et al.

Large Language Models (LLMs) represent a significant advancement in artificial intelligence, finding applications across various domains. However, their reliance on massive internet-sourced datasets for training brings notable privacy issues, which are exacerbated in critical domains (e.g., healthcare). Moreover, certain application-specific scenarios may require fine-tuning these models on private data. This survey critically examines the privacy threats associated with LLMs, emphasizing the potential for these models to memorize and inadvertently reveal sensitive information. We explore current threats by reviewing privacy attacks on LLMs and propose comprehensive solutions for integrating privacy mechanisms throughout the entire learning pipeline. These solutions range from anonymizing training datasets to implementing differential privacy during training or inference and machine unlearning after training. Our comprehensive review of existing literature highlights ongoing challenges, available tools, and future directions for preserving privacy in LLMs. This work aims to guide the development of more secure and trustworthy AI systems by providing a thorough understanding of privacy preservation methods and their effectiveness in mitigating risks.

Francesco Verdini, Pierfrancesco Melucci, Stefano Perna et al.

The remarkable performance achieved by Large Language Models (LLM) has driven research efforts to leverage them for a wide range of tasks and input modalities. In speech-to-text (S2T) tasks, the emerging solution consists of projecting the output of the encoder of a Speech Foundational Model (SFM) into the LLM embedding space through an adapter module. However, no work has yet investigated how much the downstream-task performance depends on each component (SFM, adapter, LLM) nor whether the best design of the adapter depends on the chosen SFM and LLM. To fill this gap, we evaluate the combination of 5 adapter modules, 2 LLMs (Mistral and Llama), and 2 SFMs (Whisper and SeamlessM4T) on two widespread S2T tasks, namely Automatic Speech Recognition and Speech Translation. Our results demonstrate that the SFM plays a pivotal role in downstream performance, while the adapter choice has moderate impact and depends on the SFM and LLM.

Marco Miccheli, Andrej Leban, Andrea Tacchella et al.

Translation Quality Assessment (TQA) is a process conducted by human translators and is widely used, both for estimating the performance of (increasingly used) Machine Translation, and for finding an agreement between translation providers and their customers. While translation scholars are aware of the importance of having a reliable way to conduct the TQA process, it seems that there is limited literature that tackles the issue of reliability with a quantitative approach. In this work, we consider the TQA as a complex process from the point of view of physics of complex systems and approach the reliability issue from the Bayesian paradigm. Using a dataset of translation quality evaluations (in the form of error annotations), produced entirely by the Professional Translation Service Provider Translated SRL, we compare two Bayesian models that parameterise the following features involved in the TQA process: the translation difficulty, the characteristics of the translators involved in producing the translation, and of those assessing its quality - the reviewers. We validate the models in an unsupervised setting and show that it is possible to get meaningful insights into translators even with just one review per translation; subsequently, we extract information like translators' skills and reviewers' strictness, as well as their consistency in their respective roles. Using this, we show that the reliability of reviewers cannot be taken for granted even in the case of expert translators: a translator's expertise can induce a cognitive bias when reviewing a translation produced by another translator. The most expert translators, however, are characterised by the highest level of consistency, both in translating and in assessing the translation quality.

Michele Miranda, Xinlan Yan, Nishant Mishra et al.

Protecting patient privacy in clinical narratives is essential for enabling secondary use of healthcare data under regulations such as GDPR and HIPAA. While manual de-identification remains the gold standard, it is costly and slow, motivating the need for automated methods that combine privacy guarantees with high utility. Most automated text de-identification pipelines employed named entity recognition (NER) to identify protected entities for redaction. Although methods based on differential privacy (DP) provide formal privacy guarantees, more recently also large language models (LLMs) are increasingly used for text de-identification in the clinical domain. In this work, we present the first comparative study of DP, NER, and LLMs for Dutch clinical text de-identification. We investigate these methods separately as well as hybrid strategies that apply NER or LLM preprocessing prior to DP, and assess performance in terms of privacy leakage and extrinsic evaluation (entity and relation classification). We show that DP mechanisms alone degrade utility substantially, but combining them with linguistic preprocessing, especially LLM-based redaction, significantly improves the privacy-utility trade-off.

Àlex R. Atrio, Antonio Lopez, Jino Rohit et al.

We introduce Earth Virtual Expert (EVE), the first open-source, end-to-end initiative for developing and deploying domain-specialized LLMs for Earth Intelligence. At its core is EVE-Instruct, a domain-adapted 24B model built on Mistral Small 3.2 and optimized for reasoning and question answering. On newly constructed Earth Observation and Earth Sciences benchmarks, it outperforms comparable models while preserving general capabilities. We release curated training corpora and the first systematic domain-specific evaluation benchmarks, covering MCQA, open-ended QA, and factuality. EVE further integrates RAG and a hallucination-detection pipeline into a production system deployed via API and GUI, supporting 350 pilot users so far. All models, datasets, and code are ready to be released under open licenses as contributions to our field at huggingface.co/eve-esa and github.com/eve-esa.

Branislav Gerazov, Marcello Politi, Sébastien Bratières

We explore the performance of several state-of-the-art automatic speech recognition (ASR) models on a large-scale Arabic speech dataset, the SADA (Saudi Audio Dataset for Arabic), which contains 668 hours of high-quality audio from Saudi television shows. The dataset includes multiple dialects and environments, specifically a noisy subset that makes it particularly challenging for ASR. We evaluate the performance of the models on the SADA test set, and we explore the impact of fine-tuning, language models, as well as noise and denoising on their performance. We find that the best performing model is the MMS 1B model finetuned on SADA with a 4-gram language model that achieves a WER of 40.9\% and a CER of 17.6\% on the SADA test clean set.