CST is a lightweight risk triage system that provides moderators with source‑grounded evidence for web discussions. It first assigns risk scores using character patterns and a multilingual model, then groups related messages and selects key excerpts to show, linking them back to the original thread. Across multiple languages and datasets, CST outperforms fully fine‑tuned models in ranking harmful messages and reduces moderator review time by at least 27% while maintaining high accuracy.
arXiv:2609.13579v1 Announce Type: new Abstract: Safety research often focuses on model-generated harms, but users may also direct hostility, coercion, and adversarial pressure at models. Understandin...
Safety-Flag is a unified benchmark that consolidates seven popular safety datasets into a single balanced flag/do‑not‑flag protocol, providing item‑level decisions and confidence scores for multiple large language models and dedicated guards. The benchmark evaluates moderator reliability across three dimensions—error direction, probability calibration, and confidence‑based error ranking—revealing that aggregate accuracy masks significant differences, such as one model flagging 85% of benign content while another misses 54% of harmful content. The study shows that general‑purpose models are overconfident, but temperature tuning can substantially improve calibration, and confidence‑based abstention can reduce selective risk, though performance varies with how well confidence ranks errors.
The paper presents a red‑teaming framework to evaluate how open‑source large language models (LLMs) can be used for political influence, focusing on models that can be deployed locally without weight modifications. It measures LLM Overton Windows (OWs) to assess expression fidelity of sampled political opinions and examines how natural‑language jailbreaks alter these scores. The study evaluates over 30 LLMs from 10 families across five countries, finding that left‑leaning content is expressed with higher fidelity, larger models tend to have lower mean OW scores, and jailbreak effectiveness varies significantly between model families.
arXiv:2501. 14940v4 Announce Type: replace-cross Abstract: Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption.
The paper examines how automated safety systems unevenly flag harmful content, disproportionately labeling African American English as offensive and misclassifying content from drag queens and white nationalists. It identifies seven failure modes rooted in vocabulary-based harm detection and proposes an evaluation protocol that highlights gaps in performance across languages. Applied to five open systems in eleven languages, the study finds that none achieve native-speaker accuracy, and higher capability often increases discrimination, widening the disparity between best- and worst-served language communities.
The paper introduces the SAST-IR framework to evaluate large language models’ robustness against persuasion attacks in a memory‑less setting, revealing a flaw called "Refusal Inertia" that masks true vulnerability. Using the CP‑Agent and a custom CounterFact‑Strict dataset, the authors demonstrate that simple, diverse attack strategies achieve a 96% success rate, while complex attacks often trigger defensive compliance. The study highlights severe brittleness in current state‑of‑the‑art models when deprived of conversation history.
Large language models (LLMs) increasingly support science, but they can also convert hazardous scientific knowledge into actionable misuse guidance. Existing benchmarks often rely on templated queries disconnected from real-world hazards, and employ LLM-as-a-Judge paradigms without domain grounding.
arXiv:2608. 14577v1 Announce Type: cross Abstract: Frontier large language models (LLMs) safety evaluation has largely treated harmful generation as an attack outcome rather than as an object of analysis.
arXiv:2605. 17173v2 Announce Type: replace-cross Abstract: Large language models exhibit safety degradation in non-English languages.
The paper introduces DistressBench, a benchmark comprising 718 English clinician-authored conversations covering 23 self-harm subcategories, each evaluated against a weighted rubric across seven crisis-support dimensions. Models are scored by the fraction of criteria they satisfy, revealing that the best models meet 88.4% of weighted criteria while clinicians achieve 98.7%. The study highlights gaps in crisis response, notably that 35.3% of conversations where crisis is recognized fail to route users to human help, underscoring the need to evaluate models on supportive care rather than mere refusal.
The study introduces a unified method for measuring rhetorical distortion across six social media platforms and two languages, distinguishing between creator posts and their replies. It identifies five distortion dimensions—Significance Inflation, Anxiety Manufacturing, Novelty Claims, Loaded Language, and Temporal Distortion—using a three‑tier pipeline that combines lexical rules, negative filters, and LLM verification. Results show that distortion is far more strongly linked to the platform than content type, with X/Twitter exhibiting the highest distortion scores, and that the audience layer (replies) carries significantly more distortion than creator posts.
arXiv:2608. 13695v1 Announce Type: cross Abstract: Large language model providers routinely cite multilingual safety benchmarks spanning a dozen or more languages as evidence that their models are safe for non-English-speaking users.