Toxic language includes content that is offensive, abusive, or that promotes harm. Progress in preventing toxic output from large language models (LLMs) is hampered by inconsistent definitions of toxicity. We introduce TRuST, a large-scale dataset that unifies and expands prior resources through a carefully synthesized definition of toxicity, and corresponding annotation scheme. It consists of ∼300k annotations, with high-quality human annotation on ∼11k. To ensure high-quality, we designed a rigorous, multi-stage human annotation process, and evaluated the diversity of the annotators. Then we benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, identification of the target group, and of toxic words. Our results indicate that fine-tuned PLMs outperform LLMs on the three tasks, and that current reasoning models do not reliably improve performance. TRuST constitutes one of the most comprehensive resources for evaluating and mitigating LLM toxicity, and other research in socially-aware and safer language technologies.

Existing hate speech detection models are often opaque and rely on surface-level lexical cues, which makes them vulnerable to spurious correlations and limits robustness, interpretability and cultural contextualization. We propose Supervised Moral Rationale Attention (SMRA), the first self-explaining hate speech detection framework to incorporate moral rationales as direct supervision for attention alignment. Based on Moral Foundations Theory, SMRA aligns token-level attention with expert-annotated moral rationales, guiding models to attend to morally salient spans. Unlike prior rationale-supervised or post-hoc approaches, SMRA integrates moral rationale supervision directly into the training objective, producing inherently interpretable and contextualized explanations. To support our framework, we also introduce HateBRMoralXplain, a Brazilian Portuguese benchmark dataset annotated with hate labels, moral categories, token-level moral rationales, and socio-political metadata. Across binary hate speech detection and multi-label moral sentiment classification, SMRA consistently improves performance while enhancing both faithful and plausible explanations. Although explanations become more concise, sufficiency decreases, indicating more compact and informative rationales. Fairness remains stable, suggesting that improvements in explanation quality do not introduce significant bias trade-offs.

Large language models (LLMs) have become ubiquitous, thus it is important to understand their risks and limitations, such as their propensity to generate harmful output. This includes smaller LLMs, which are important for settings with constrained compute resources, such as edge devices. Detection of LLM harm typically requires human annotation, which is expensive to collect. This work studies two questions: How do smaller LLMs rank regarding generation of harmful content? How well can larger LLMs annotate harmfulness? We prompt three small LLMs to elicit harmful content of various types, such as discriminatory language, offensive content, privacy invasion, or negative influence, and collect human rankings of their outputs. Then, we compare harm annotation from three state-of-the-art large LLMs with each other and with humans. We find that the smaller models differ with respect to harmfulness. We also find that large LLMs show low to moderate agreement with humans.

LLM (large language model) users of hosted providers commonly notice that outputs can vary for the same inputs under settings expected to be deterministic. While it is difficult to get exact statistics, recent reports on specialty news sites and discussion boards suggest that among users in all communities, the majority of LLM usage today is through cloud-based APIs. Yet the questions of how pervasive non- determinism is, and how much it affects perfor- mance results, have not to our knowledge been systematically investigated. We apply five API- based LLMs configured to be deterministic to eight diverse tasks across 10 runs. Experiments reveal accuracy variations of up to 15% across runs, with a gap of up to 70% between best pos- sible performance and worst possible perfor- mance. No LLM consistently delivers the same outputs or accuracies, regardless of task. We speculate about the sources of non-determinism such as input buffer packing across multiple jobs. To better quantify our observations, we introduce metrics focused on quantifying de- terminism, TARr@N for the total agreement rate at N runs over raw output, and TARa@N for total agreement rate of parsed-out answers. Our code and data will be publicly available at https://github.com/Anonymous.