LLMs Fail to Recognize Quantization-Induced Performance Degradation
October 6, 2026
Probing shows that LLMs cannot self-report degradation caused by quantization, despite internal representations containing clear fingerprints of the quantization method. While a shared LoRA can identify these states, the models themselves lack self-awareness of their computational substrate.
HOW THIS AFFECTS YOU
●
builderDo not rely on an LLM's self-assessment to verify if it is running in a degraded, quantized state.
●
researcherThis identifies a fundamental gap in LLM self-awareness and model monitoring.