Breaking01 / AI
Classifier outperforms model at detecting perturbations
Researchers cautioned that identifying a perturbation's location does not imply broad access to a model's internal computation.

Researchers reported 98.3% detection accuracy for a classifier during pretraining under specific perturbation and readout conditions in OLMo 3 32B, in an arXiv paper published on 13 Nov 2026.
According to the paper, the researchers trained a linear classifier, a simple tool for sorting signals, to identify the perturbed sentence from the model's internal signals. They then compared its predictions with the model's own answers.
The paper-reported accuracy applies to a specific experimental setup. Under the same conditions, the model's answers without example prompts reached 11.1% accuracy, equivalent to chance.
At the publicly released base checkpoint, the paper reports classifier accuracy of 99.1% and model answer accuracy of 89.6%. A checkpoint is a model saved at a particular stage of training.
The researchers also reported that perturbation locations could be decoded in randomly initialised models. They said this showed that model architecture can retain location traces without pretraining, and cautioned that successful location reporting does not imply broader access to internal computation.
The paper's abstract does not provide a score measuring how faithfully the model explains its reasoning.