Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate
This paper investigates whether LLMs can recognize degradation in their own computational substrate (i.e., quantization). It demonstrates that while existing LLMs cannot identify their quantization status through self-reporting, distinguishable differences exist within their internal states, which can be read out through additional training. However, this readout merely maps these internal state differences to answer labels, failing to generalize to other quantization techniques.
