toxicity

This paper presents the first comparative evaluation of training-free methods for personalizing toxicity sensitivity in language models at inference time, showing that all methods reduce alignment error by 28-47% but reveal a trade-off between alignment, personalization, and language quality.

ToxiREX is a new multilingual dataset of Reddit comments annotated for implicit toxicity using a toxic reasoning schema, covering six languages and multiple events.

TriEval is a new pipeline for evaluating LLMs across bias, toxicity, and truthfulness simultaneously, designed to be resource-efficient and run on standard laptops. It has been tested on Llama 3 8B, Mistral 7B, Gemma 2 9B, and Claude Haiku, and is released as open source.

This paper investigates how toxic lexical perturbations in prompts reduce the factual accuracy and increase uncertainty of LLMs, and uses attribution-graph analyses to trace internal changes. It finds that increasing toxicity amplifies perturbation-sensitive variant nodes while core reasoning nodes remain invariant.

The paper argues that unlearning in LLMs should be goal-dependent, proposing a cosine-based meta-learned variant of RMU for dangerous knowledge and a multi-layer objective with probe directions for toxicity, achieving strong results across four 7-8B models.

This paper uses a pre-trained LLM with zero-shot classification to analyze approximately 20 million Twitch chat messages across seven game genres, finding that 2.4% of messages are toxic, with MOBA games having the highest rate (3.2%) and sports games the lowest (2%). The study also identifies significant differences in toxicity distributions across individual games within the same genre.

This paper identifies and studies 'memory laundering' in LLM agents, where toxic or adversarial context compressed into memory summaries evades standard toxicity detectors while still influencing future generations. It introduces the sub-threshold propagation gap (SPG) to measure hidden downstream influence and shows that sanitizing toxic state before summarization is more effective than post-hoc cleaning.

This replication study evaluates DExperts for mitigating toxicity in LLMs, finding near-perfect safety against explicit toxicity but reduced effectiveness against implicit hate speech and a significant latency trade-off.

A Stanford study analyzing billions of social media posts reveals that only ~3% of users generate severely toxic content, but engagement-driven algorithms disproportionately amplify this minority, distorting public perception and driving self-censorship among the majority.