arXiv subjects

Pedro O. S. Vaz-de-Melo

Publications and source records attributed to Pedro O. S. Vaz-de-Melo.

3 recordsLinked to original sources

arXiv2026

Ideological discrepancy between Facebook pages and shared news is associated with reaction homogeneity

Cross-cutting political content is often expected to mitigate polarization, yet encounters across ideological lines do not necessarily foster more constructive engagement. We analyzed 4,584 posts from 1,243 Facebook pages sharing 761 political news links in a dataset centered on Brazil's 2018 presidential election, and introduced Bias Discrepancy, which measures how far a shared link lies from the page's ideological sharing repertoire. We examined how Bias Discrepancy related to Content Reaction Positivity (how positively the same link was received in other posts), the Reaction Homogeneity Index (whether a post's own reactions were one-sided or mixed), and the toxicity of the page's accompanying text. Reaction homogeneity was lowest at both extremes of Bias Discrepancy, and posts furthest from the page's repertoire had the highest average toxicity. In a multivariable model, greater Bias Discrepancy was associated with lower reaction homogeneity, with a magnitude similar to the link's content bias and toxicity and smaller than Content Reaction Positivity. When a page shared news that departed from its repertoire, its accompanying text tended to be more toxic and its audience's reactions more mixed.

cs.SI↗

arXiv2026

Widespread Gender and Pronoun Bias in Moral Judgments Across LLMs

Large language models (LLMs) are increasingly used to assess moral or ethical statements, yet their judgments may reflect social and linguistic biases. This work presents a controlled, sentence-level study of how grammatical person, number, and gender markers influence LLM moral classifications of fairness. Starting from 550 balanced base sentences from the ETHICS dataset, we generated 26 counterfactual variants per item, systematically varying pronouns and demographic markers to yield 14,850 semantically equivalent sentences. We evaluated six model families (Grok, GPT, LLaMA, Gemma, DeepSeek, and Mistral), and measured fairness judgments and inter-group disparities using Statistical Parity Difference (SPD). Results show statistically significant biases: sentences written in the singular form and third person are more often judged as "fair'', while those in the second person are penalized. Gender markers produce the strongest effects, with non-binary subjects consistently favored and male subjects disfavored. We conjecture that these patterns reflect distributional and alignment biases learned during training, emphasizing the need for targeted fairness interventions in moral LLM applications.

cs.CL↗

arXiv2022

Development and Validation of an AI-Driven Model for the La Rance Tidal Barrage: A Generalisable Case Study

In this work, an AI-Driven (autonomous) model representation of the La Rance tidal barrage was developed using novel parametrisation and Deep Reinforcement Learning (DRL) techniques. Our model results were validated with experimental measurements, yielding the first Tidal Range Structure (TRS) model validated against a constructed tidal barrage and made available to academics. In order to proper model La Rance, parametrisation methodologies were developed for simulating (i) turbines (in pumping and power generation modes), (ii) transition ramp functions (for opening and closing hydraulic structures) and (iii) equivalent lagoon wetted area. Furthermore, an updated DRL method was implemented for optimising the operation of the hydraulic structures that compose La Rance. The achieved objective of this work was to verify the capabilities of an AI-Driven TRS model to appropriately predict (i) turbine power and (ii) lagoon water level variations. In addition, the observed operational strategy and yearly energy output of our AI-Driven model appeared to be comparable with those reported for the La Rance tidal barrage. The outcomes of this work (developed methodologies and DRL implementations) are generalisable and can be applied to other TRS projects. Furthermore, this work provided insights which allow for more realistic simulation of TRS operation, enabled through our AI-Driven model.

cs.LG↗