The Geometry of Harmful Intent: Training-Free Anomaly Detection via Angular Deviation in LLM Residual Streams

1. 1Compute the leading principal component of activations for 200 safe normative prompts at a target layer
2. 2Characterise new prompts by their radial deviation angle from the reference direction
3. 3Calculate the anomaly score as the negative log-likelihood of the deviation angle
4. 4Use the anomaly score to detect harmful prompts

AI researchers and engineers working on LLMs can benefit from this method to detect harmful prompts, and it can be used by product managers to improve the safety of their AI-powered products

💡 Angular deviation in residual streams can be used to detect harmful prompts in LLMs without training

💡 Detect harmful prompts in LLMs without training! LatentBiopsy uses angular deviation in residual streams