Virtue AI Research

Research Insights

From academic research to production-grade AI safety.

HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding.

Differentially Private Synthetic Data via Foundation Model APIs 2:

HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal.

Effects of Exponential Gaussian Distribution on (Double Sampling) Randomized Smoothing.

Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression.

Rob-FCP: Certifiably Byzantine-Robust Federated Conformal Prediction.

RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content.

C-RAG: Certified Generation Risks for Retrieval-Augmented Language Models

Fine-tuning aligned language models compromises safety, even when users do not intend to!