Keep Calm and Avoid Harmful Content: Concept Alignment and Latent Manipulation Towards Safer Answers

CALM addresses a real vulnerability: Large language models can be tricked into generating harmful content through cleverly crafted prompts (jailbreak attacks), even when safety features are built in—this paper proposes a practical solution that works at the moment the model generates responses.

No retraining required: Unlike many safety approaches, CALM modifies how the model processes information during inference without requiring expensive retraining or new training data, making it a practical addition to existing AI systems.

The method works by removing harmful "directions": The technique identifies and suppresses specific patterns in the model's internal representations associated with harmful content, similar to removing unwanted features from an image, while keeping the model's useful capabilities intact.

Lightweight computational cost: CALM adds only minimal computational overhead when generating responses, meaning it can be deployed without significantly slowing down AI systems or increasing operational costs.

Outperforms existing safeguards: Experiments demonstrate that CALM reduces harmful outputs more effectively than baseline safety methods in most tested scenarios, suggesting it could meaningfully improve AI safety without sacrificing model usefulness.