Research shows compressed models can hide harmful behaviors activated only after INT8 or 4-bit quantization, creating a gap between validation and deployment.
Post-training compression is not neutral. A paper shows behavioral equivalence classes collapsing distinct parameters and enabling backdoors that activate only after INT8 or 4-bit quantization. Tests on multilingual models report up to 85.02% friend-...
A new study formalizes the validation-deployment gap: an LLM can pass full-precision checks and activate harmful behavior only after INT8 or 4-bit quantization. In tests, tactical translation moves from zero friend-foe corruption at repaired FP16 to ...
