When an LLM reads a document, it ingests both the core factual data and the rhetorical style, bias, and intent of the author as a single, entangled concept. KP-IDP introduces a framework to mathematically separate these two elements during pre-training.
Potential State
The pure, objective factual baseline of the information. Stripped of emotion, persuasion, or subjective framing. This forms the bedrock of the model's knowledge graph.
Kinetic State
The rhetorical energy, tone, and persuasive intent. By disentangling this, the model understands *how* something is being said without blindly adopting that perspective as absolute truth.
By actively disentangling information into Potential and Kinetic components, we can train foundation models that possess vast knowledge without becoming ideologically captured by the tonal biases inherent in human-generated training data.