Principal Component Projection Improves Llama-3.1-8B Deception Detection Generalization
September 4, 2026
Projecting Llama-3.1-8B-Instruct activations onto a subset of training-distribution principal components enables out-of-distribution deception detection. Using an LLM judge to select the highest-scoring transferable components closed the baseline-to-oracle gap by 78% on Insider Trading Report tasks.
HOW THIS AFFECTS YOU
●
researcherYou can improve linear probe robustness by restricting analysis to transferable subspace directions.