Topic
arXiv
-
News AI
Researchers find inference-time model changes invalidate conformal safety bounds in LLMs
Auditing conformal prediction wrappers across language models revealed that silent weight quantization and routing shifts destroy mathematical uncertainty guarantees.