BeaverTails

A reproducibility study of latent-space safety probes found that lightweight MLP probes trained on final-layer activations of LLaMA-3.1-8B detect harmful prompts with F1 scores within 0.37 percentage …

A new approach called DataRx treats LLM safety as a missingness problem by analyzing hidden representations to sample only the most needed safety data, reducing attack success rates from 59.23% to 13.…

A new method called Geometry-Guided Constraint Learning improves LLM safety classification by using sparse autoencoder (SAE) feature extraction, achieving 96-99% accuracy per category on the BeaverTai…

A comprehensive evaluation of 14 open-source safety guard models on a benchmark of 79,331 samples found that Qwen Guard, a 4-billion-parameter model, achieved the highest recall at 83.97%, while large…