JAM (Just A Move) is a novel framework that enables interpretable and responsible control of language models by manipulating their internal latent representations based on causal reasoning.

• Provides a cause-effect analysis within the latent space of LLMs
• Enables fine-grained control over text generation characteristics
• Demonstrates significant toxicity reduction while maintaining text quality
• Offers a transparent approach to responsible AI text generation

Security Impact: JAM addresses critical concerns about harmful content generation by providing a systematic way to reduce toxic outputs without sacrificing performance, making AI systems safer for deployment in sensitive environments.

JAM: Controllable and Responsible Text Generation via Causal Reasoning and Latent Vector Manipulation