Problem: Probes that predict concepts from language models are usually judged by accuracy. But accuracy can hide reliance on correlated but irrelevant attributes, like gender for occupation. So a high-accuracy probe may not measure what we intend.
What was done: The authors compared probes with equal accuracy across four settings. They measured sensitivity to nuisances and found several-fold differences. They also developed diagnostics that require only nuisance labels to select better probes.
Why it matters: Accuracy alone is an unreliable guide; equally accurate probes can have very different focus. Their proposed diagnostics improve probe selection, making concept measurements more trustworthy.