Research
Linear bias probes fail LLMs at specific sample sizes
Linear bias probes in LLM representations fail predictably when Fisher estimation error exceeds the eigengap, a spectral condition that makes probe instability diagnosable before deployment.