Back to AI Research

AI Research

Where Should Knowledge Enter? A Layered Framework f... | AI Research

Key Takeaways

  • Multimodal generative models, such as those that create images from text, are often fluent but struggle to follow strict rules, safety guidelines, or domain-...
  • Multimodal generative models produce fluent outputs but remain unreliable when generation must respect structured, domain-specific, or safety-critical knowledge.
  • We argue that knowledge infusion in iterative generative models is fundamentally anintervention-layer problem.
  • This maps to four intervention layers: surface, trajectory, latent, and parametric infusion.
  • We instantiate the framework in diffusion models, map representative methods to all four layers, and derive design principles for multi-layer composition.
Paper AbstractExpand

Multimodal generative models produce fluent outputs but remain unreliable when generation must respect structured, domain-specific, or safety-critical knowledge. Existing methods incorporate knowledge through mechanisms such as prompt augmentation, guidance, latent editing, or fine-tuning, yet they are typically categorized by technique rather than by the component of the generative process they modify. We argue that knowledge infusion in iterative generative models is fundamentally anintervention-layer problem. Since thegenerative process unfolds as a trajectory of internal states, knowledge can act on four structurally distinct components of this process: the input/output boundary, the transition function, the intermediate state, and the model parameters. This maps to four intervention layers: surface, trajectory, latent, and parametric infusion. We instantiate the framework in diffusion models, map representative methods to all four layers, and derive design principles for multi-layer composition. In a controlled safety-alignment experiment using a multimodal knowledge graph with two diffusion backbones, we implement three of the four layers cumulatively, surface (input-side and output-side) and trajectory--latent (mid-generation). We show empirically that each additional layer addresses failure classes that prior layers cannot reach, reducing knowledge-violating outputs by 70.97% compared to vanilla generation and empirically confirming the framework's complementarity prediction.

Multimodal generative models, such as those that create images from text, are often fluent but struggle to follow strict rules, safety guidelines, or domain-specific facts. This paper proposes a new framework to solve this by treating "knowledge infusion"—the process of injecting external information into a model—as an intervention-layer problem. Instead of viewing knowledge integration as a collection of disconnected techniques, the authors categorize these methods based on which specific part of the generative process they modify, providing a roadmap for how to best align AI outputs with reliable knowledge.

The Four Layers of Generation

The researchers observe that iterative generative models, like diffusion models, produce outputs through a trajectory of internal states. They identify four distinct points along this trajectory where external knowledge can be applied:

  • Surface Infusion: Acts at the boundaries of the process by modifying the input prompt or post-processing the final output.

  • Trajectory Infusion: Modifies the "transition function" during the generation process, essentially steering the model’s decision-making at every step.

  • Latent Infusion: Directly alters the intermediate hidden states of the model while it is in the middle of generating an output.

  • Parametric Infusion: Changes the model’s actual weights or architecture, permanently embedding knowledge into the system before it ever starts generating.

Why the Framework Matters

By mapping existing methods to these four layers, the authors clarify why certain techniques work better for specific problems. For example, surface infusion is simple and cheap but cannot fix deep-seated structural errors. Conversely, parametric infusion is powerful and permanent but requires expensive retraining. This framework helps practitioners choose the right tool for the job: if a model fails to follow a specific rule, they can now determine whether to adjust the prompt (surface), steer the sampler (trajectory), edit the internal state (latent), or retrain the model (parametric).

Empirical Results and Complementarity

To test this, the researchers conducted a safety-alignment experiment using two diffusion models. They implemented a multi-layer approach, combining surface, trajectory, and latent interventions. The results showed that these layers are complementary; each additional layer addressed types of failures that the previous layers could not reach. By stacking these interventions, the team reduced knowledge-violating outputs by 70.97% compared to a standard, "vanilla" model. This confirms that no single layer is a "silver bullet," and that a principled, layered approach is necessary for high-fidelity, knowledge-consistent generation.

Key Design Principles

The authors suggest that when building these systems, developers should focus on three core principles: matching the intervention layer to the specific class of failure, composing multiple layers to ensure full coverage of potential errors, and carefully managing potential interference between different layers. By following this structure, developers can create more reliable AI systems that remain faithful to structured knowledge without sacrificing the creative fluency of the underlying model.

Comments (0)

No comments yet

Be the first to share your thoughts!