Evaluating the Realism of LLM-powered Social Agents: A Case Study of Reactions to Spanish Online News

LLM-powered social agents are increasingly used to simulate online social behavior, yet their realism remains difficult to validate. Existing work has largely relied on general-purpose benchmarks, while less attention has been paid to short, reactive discourse such as audience replies to online news. In this paper, we evaluate whether LLM-generated reactions to Spanish online news reproduce measurable properties of real audience discourse. Using the Hatemedia dataset, we pair 5,631 news items with 58,555 real audience reactions, and generate a matched synthetic dataset using five LLMs under a shared experimental setting. We compare real and synthetic reactions across three dimensions: hate speech, sentiment, and semantic alignment, considering both off-the-shelf and fine-tuned generation. Results show that off-the-shelf models are poor proxies for real audience reactions: they strongly underproduce hate speech, introduce model-specific sentiment biases, and remain distributionally distant from human replies. Fine-tuning improves fidelity unevenly. Qwen3 provides the most balanced approximation, while Mistral7B achieves the strongest sentiment and semantic alignment but overshoots hate prevalence. Plausible synthetic replies do not necessarily reproduce the distributional properties of public discourse.

A Big Data Architecture for Early Identification and Categorization of Dark Web Sites

The dark web has become notorious for its association with illicit activities and there is a growing need for systems to automate the monitoring of this space. This paper proposes an end-to-end scalable architecture for the early identification of new Tor sites and the daily analysis of their content. The solution is built using an Open Source Big Data stack for data serving with Kubernetes, Kafka, Kubeflow, and MinIO, continuously discovering onion addresses in different sources (threat intelligence, code repositories, web-Tor gateways, and Tor repositories), downloading the HTML from Tor and deduplicating the content using MinHash LSH, and categorizing with the BERTopic modeling (SBERT embedding, UMAP dimensionality reduction, HDBSCAN document clustering and c-TF-IDF topic keywords). In 93 days, the system identified 80,049 onion services and characterized 90% of them, addressing the challenge of Tor volatility. A disproportionate amount of repeated content is found, with only 6.1% unique sites. From the HTML files of the dark sites, 31 different low-topics are extracted, manually labeled, and grouped into 11 high-level topics. The five most popular included sexual and violent content, repositories, search engines, carding, cryptocurrencies, and marketplaces. During the experiments, we identified 14 sites with 13,946 clones that shared a suspiciously similar mirroring rate per day, suggesting an extensive common phishing network. Among the related works, this study is the most representative characterization of onion services based on topics to date.