This research introduces a framework that makes LLM embeddings interpretable and controllable for classification tasks, addressing regulatory compliance and model generalization challenges.

• Proposes a self-regularization mechanism that reveals unintended features in LLM embeddings
• Enables selective removal of sensitive or task-irrelevant features from classification models
• Demonstrated improvements in toxic content detection and other security applications
• Achieves better generalization by focusing only on task-relevant features

For security teams, this approach offers a critical advancement in creating more transparent, compliant AI systems where unwanted biases or problematic features can be identified and removed.

Self-Regularization with Latent Space Explanations for Controllable LLM-based Classification