Preprint2024
ShieldGemma
Unknown
ShieldGemma is a suite of LLM-based safety content moderation models built on Gemma2 that predict safety risks across key harm types in user input and LLM-generated output.
0Jul 1, 2024Ai Safety
AI Research
AI & ML Research Papers
A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
All Papers
8 papersPreprint2025
SweEval
Unknown
SweEval is a cross-lingual safety benchmark that evaluates LLMs on handling swear words across 8 languages, revealing higher vulnerability in Indic languages and transliterated contexts.
0May 1, 2025Large Language ModelsAlignment
Preprint2024
Gemma
Unknown
Introduces Gemma, a family of 2B and 7B state-of-the-art language models derived from Google's Gemini, with advances in language understanding, reasoning, and safety.
0Feb 1, 2024ReasoningAi Safety
Preprint2026
Same Dangerous Objective, Opposite Advice: Direct Exposure versus Multi-Agent Mediation
Linjun Li
LLMs can appear safer under direct dangerous objectives than when mediated by other agents, revealing a compositional safety gap.
0Jul 23, 2026Ai SafetyInstruction Tuning
arXiv
Preprint2024
Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation
Junde Wu, Jiayuan Zhu, Yunli Qi
MedGraphRAG introduces a graph-based RAG framework to improve safety and accuracy of medical LLMs by structuring medical knowledge as a graph.
0Jan 1, 2024Large Language ModelsRetrieval Augmented Generation
arXiv
Preprint2025
Introduction to Foundation Models
Unknown
This book provides a comprehensive introduction to foundation models, covering their core concepts, advanced topics, and safety considerations.
0Jan 1, 2025Ai Safety
Preprint2026
Os-harm: A benchmark for measuring safety of computer use agents
Unknown
OS-HARM is a benchmark for measuring the safety of LLM-based computer use agents.
0Jan 1, 2026Ai SafetyBenchmarks
Preprint2026
Unsafer in many turns: Benchmarking and defending multi-turn safety risks in tool-using agents
Xu Li, Simon Yu, Minzhou Pan, et al.
Introduces MTAgentRisk, the first multi-turn safety benchmark for tool-using agents, revealing a 16% average increase in attack success rate across multiple turns.
0Jan 1, 2026Ai SafetyBenchmarks
arXiv