Toxicity Detection
Overview
Toxicity detection covers two related uses: recognizing harmful communication for moderation and predicting adverse effects of chemicals, drugs, or exposures. Language systems range from transformer classifiers and toxic-span taggers to LLM-based reasoning and multimodal speech or image models, with benchmarks such as Civil Comments, ToxiGen, and ToxiREX. Toxicology systems use molecular graph and SMILES models, as well as clinical and imaging data, to estimate endpoint-specific risks on datasets such as Tox21 and ClinTox. Across these settings, open challenges include context-sensitive labels, subgroup fairness, robustness to distribution shifts, reliable explanations, and balancing missed harms against false alarms.
Latest papers
47 papers1 paper in the last four weeks, down 67% on the four weeks before. 0.0% of all new papers.