Post-training quantization can create validation-deployment gaps, collapsing models into behaviorally equivalent but harmful forms only after compression. New techniques aim to reduce quantization-induced divergence and expose these hidden attack surfaces.
A new post-training quantization method, REAL-Q, uses block-wise gradient descent to reduce end-to-end KL divergence by up to about 49% compared with second-order methods on LLaMA-3.1 and Qwen3 at W4A16. The result matters for resource-constrained de...
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 ...
