Tag
The paper finds that bitsandbytes INT4 quantization significantly amplifies proactive interference in LLMs, degrading accuracy in contexts with repeated overwrites and highlighting deployment risks for semantically dense applications.