Building Self-Awareness of LLMs Over Weight Quantization

Large Language Models (LLMs) are increasingly deployed in resource-constrained settings using lossy compression techniques such as weight quantization. While effective for reducing memory and computational demands, these techniques often degrade performance. This paper investigates a novel question: Can an LLM recognize that it has been lossy-compressed? We propose a probing framework that combines few-shot prompting with introspective queries to assess whether a model can identify its own quantization state. Experiments across multiple LLMs, including Qwen, Mistral, LLaMA, and GPT-3.5, demonstrate that many models can accurately detect their compression status, especially when provided with reference examples. These findings introduce a new perspective on model self-awareness and have broader implications for research on transparency and interpretability in compressed LLMs.