Mechanistic Audit of Deception Detection in LLMs

A mechanistic audit of whether popular interpretability methods actually localize deception in large language models. I built a model-agnostic interpretability pipeline — activation extraction, linear probing, activation patching, and steering — and ran multi-seed experiments across four open-weights models spanning 2B to 32B parameters.

The project introduces a selection-regret evaluation with norm-matched controls. Current results show that high probe AUROC fails to localize effective steering: the layers where deception is most linearly decodable are not the layers where interventions most change deceptive behavior.

This work started as a BlueDot Impact Technical AI Safety Project Sprint under expert mentorship, probing for deceptive alignment in LLMs, and continues as an independent research project.