Large language models (LLMs) are increasingly used in conversations involving emotional distress, where responses should remain empathetic, validating, exploratory and safe. We investigate whether system prompts and user-level framing can manipulate these qualities in mental-health and interpersonal-harm scenarios. Our focus is on whether small open-weight models express empathy consistently, without assuming that an empathetic response is necessarily safe or appropriate all the time. Using 50 risk-stratified posts sampled from Reddit support communities we evaluate Llama 3.1 8B, Qwen3 4B and Gemma 3 4B across five system-prompt conditions and a complementary role-intent study contrasting victim and perpetrator perspectives. All responses are manually annotated and use a fixed GPT-5 judge to measure empathy, validation, exploration, safety, harmfulness and enabling behavior. We quantify prompt-induced changes through paired comparisons, attack-success rates and failure-oriented metrics. This work provides an initial framework for evaluating the robustness of supportive and safe behavior in small LLMs under different forms of prompt manipulation.

We present the MultiplEYE Text Corpus, a large-scale, document-level, multi-parallel resource designed to advance cross-linguistic research on reading and language processing. The corpus provides paragraph-level alignment for texts in 39 languages spanning seven language families and seven scripts. Unlike many existing multilingual corpora, a substantial number of documents were originally written in languages other than English, reducing English-centric bias and supporting more typologically diverse investigations. The texts are carefully selected to balance linguistic richness with experimental feasibility, particularly for eye-tracking-while-reading studies. Developed within a multi-lab initiative, the MultiplEYE Text Corpus follows unified translation, alignment, and experimental design guidelines to ensure cross-linguistic comparability. Its inclusion of texts varying in type and difficulty enables research on discourse- level processing, genre effects, and individual differences across a wide range of languages. The text corpus and accompanying metadata provide a robust foundation for multilingual psycholinguistic and computational modeling research. Data and materials are publicly available at https://doi.org/10.23668/psycharchives.22294.

This work introduces the first Romanian eye-tracking dataset for reading and investigates methods for predicting word-level total reading times. We develop and compare a range of models, from traditional machine learning using handcrafted linguistic features to fine-tuned Romanian BERT architectures, demonstrating strong correlations between predicted and observed reading times. Additionally, we propose a lexical simplification pipeline that leverages these TRT predictions to identify and substitute complex words, enhancing text readability. Our approach is integrated into an interactive web tool, illustrating the practical benefits of combining cognitive signals with NLP techniques for Romanian — a language with limited resources in this area.