Open Internet by MindsNet

Dialect-Conditioned Safety Failures in MoE Models

MoE language models exhibit dialect-conditioned safety failures, particularly when refusal behavior is weakened or removed. The models respond differently to African American English Vernacular (AAVE) and Academic English (AE) prompts in safety-sensitive situations. This raises concerns about the reliability of these models in real-world deployments.

Computing & Technology, Computer Science, Machine Learning