Anastasia Giachanou

Chain-of-thought (CoT) reasoning has emerged as a powerful technique for improving the problem-solving capabilities of large language models (LLMs), particularly for tasks requiring multi-step reasoning. However, recent studies show that CoT explanations often fail to reflect the model's actual reasoning process, as models may produce coherent yet misleading justifications or modify answers without acknowledging external cues. Such discrepancies undermine the reliability of CoT-based methods for safety supervision and alignment monitoring, as models can generate plausible but deceptive rationales for incorrect answers. To better understand this limitation, we evaluate two optimization methods, Group Relative Policy Optimization (GRPO) and Direct Preference Optimization (DPO), in their ability to improve CoT faithfulness. Our experiments show that GRPO achieves higher performance than DPO in larger models, with the Qwen2.5-14B-Instruct model attaining the best results across all evaluation metrics. Both approaches exhibit positive correlations between model size and performance, but GRPO shows greater potential for improving faithfulness metrics, albeit with less stable behavior at smaller scales. These results suggest that GRPO offers a promising direction for developing more transparent and trustworthy reasoning in LLMs.

Stereotypes influence social perceptions and can escalate into discrimination and violence. While NLP research has extensively addressed gender bias and hate speech, stereotype detection remains an emerging field with significant societal implications. This work presents a survey of existing research, drawing on definitions from psychology, sociology, and philosophy. A semi-automatic literature review was conducted using Semantic Scholar, through which over 6,000 papers (published between 2000–2025) were retrieved and filtered. The analysis identifies key trends, methodologies, challenges, and future directions. The findings emphasize the potential of stereotype detection as an early-monitoring tool to prevent bias escalation and the rise of hate speech. The conclusions call for a broader, multilingual, and intersectional approach in NLP studies.

In this study, we present a new approach that combines multiple Bidirectional Encoder Representations from Transformers (BERT) architectures with a Convolutional Neural Network (CNN) framework designed for sexism detection in text at a granular level. Our method relies on the analysis and identification of the most important terms contributing to sexist content using Shapley Additive Explanations (SHAP) values. This approach involves defining a range of Sexism Scores based on both model predictions and explainability, moving beyond binary classification to provide a deeper understanding of the sexism-detection process. Additionally, it enables us to identify specific parts of a sentence and their respective contributions to this range, which can be valuable for decision makers and future research. In conclusion, this study introduces an innovative method for enhancing the clarity of large language models (LLMs), which is particularly relevant in sensitive domains such as sexism detection. The incorporation of explainability into the model represents a significant advancement in this field. The objective of our study is to bridge the gap between advanced technology and human comprehension by providing a framework for creating AI models that are both efficient and transparent. This approach could serve as a pipeline for future studies to incorporate explainability into language models.

With the explosive growth of online social media, the ancient problem of information disorders interfering with news diffusion has surfaced with a renewed intensity threatening our democracies, public health, and news outlets' credibility. Therefore, thousands of scientific papers have been published in a relatively short period, making researchers of different disciplines struggle with an information overload problem. The aim of this survey is threefold: (1) we present the results of a network-based analysis of the existing multidisciplinary literature to support the search for relevant trends and central publications; (2) we describe the main results and necessary background to attack the problem under a computational perspective; (3) we review selected contributions using network science as a unifying framework and computational linguistics as the tool to make sense of the shared content. Despite scholars working on computational linguistics and networks traditionally belong to different scientific communities, we expect that those interested in the area of fake news should be aware of crucial aspects of both disciplines.

Stance detection (SD) can be considered a special case of textual entailment recognition (TER), a generic natural language task. Modelling SD as TER may offer benefits like more training data and a more general learning scheme. In this paper, we present an initial empirical analysis of this approach. We apply it to a difficult but relevant test case where no existing labelled SD dataset is available, because this is where modelling SD as TER may be especially helpful. We also leverage measurement knowledge from social sciences to improve model performance. We discuss our findings and suggest future research directions.

Studies assessing the effects of social media use are largely based on measures of time spent on social media. In recent years, scholars increasingly ask for more insights in social media activities and content people engage with. Data Download Packages (DDPs), the archives of social media platforms that each European user has the right to download, provide a new and promising method to collect timestamped and content-based information about social media use. In this paper, we first detail the experiences and insights of a data collection of 110 Instagram DDPs gathered from 102 adolescents. We successively discuss the challenges and opportunities of collecting and analyzing DDPs to help future researchers in their consideration of whether and how to use DDPs. DDPs provide tremendous opportunities to get insight in the frequency, range, and content of social media activities, from browsing to searching and posting. Yet, collecting, processing, and analyzing DDPs is also complex and laborious, and demands numerous procedural and analytical choices and decisions.

Recent years have seen a tremendous increase in the propagation of different types of misinformation and disinformation, including among others fake news, rumours, clickbait and conspiracy theories. Misinformation involved in satire and clickbait, among others, has a different intention from disinformation. Despite many attempts by the research community, the development of technology to assist experts in detecting disinformation remains an open problem due to a number of challenges. For example, there are various biases such as confirmation bias and peer pressure that hinder users from recognizing non-credible information. In addition, fake news is intentionally written to confuse the readers, often containing a mixture of false and real information. To this end, in this editorial we present current challenges in the area of fake news identification and discuss contributions published in our special issue.