A reference-free auditing method detects planted behaviors in LLMs by identifying late-layer activation residuals that deviate from early-layer predictions.

This research presents a method for detecting hidden, planted behaviors in LLMs without requiring a reference base model. By analyzing the residuals between early-layer and late-layer activations via Ridge regression, the approach identifies where models deviate from their internal logic, effectively "ratting themselves out." Tested on Anthropic's AuditBench organisms, the method achieved AUROC scores up to 0.889, matching or exceeding reference-based benchmarks. The project also introduces a "topic funnel" and probe specificity filters to distinguish between broad RLHF-induced biases and narrow, topic-specific fine-tuning.