Researchers exploit quantization to hide malicious AI model behavior

Original: Widening the Gap: Exploiting LLM Quantization via Outlier Injection

Source: arxiv.org ↗