A neural network architecture technique in which inputs are directed to different specialized subnetworks based on their content, rather than every input passing through every parameter. Dynamic routing is the mechanism behind mixture-of-experts models, which is how many of the largest and most capable AI systems achieve high performance without proportional increases in compute costs.
Also known as adaptive routing, mixture-of-experts routing, expert gating
Dynamic routing allows different inputs to follow different paths through a neural network. In mixture-of-experts architectures, a gating network examines each input and routes it to a small subset of specialized expert subnetworks, rather than passing it through all parameters. The expert subnetworks specialize through training: some become good at processing certain input types while others become good at different ones. The routing decision is made dynamically for each input, allocating computation to the experts most relevant for each case.
The practical consequence is scale efficiency. A model with 100 billion parameters would normally require all 100 billion to be activated for every input, making inference extremely expensive. A mixture-of-experts model with 100 billion total parameters but routing that activates only 10 billion for each input delivers performance closer to a 100-billion parameter model at a compute cost closer to a 10-billion parameter model. This is why several of the most capable large language models deployed commercially use mixture-of-experts architectures.
Dynamic routing also applies in other contexts: content delivery networks use dynamic routing to direct web requests to the server best positioned to respond, and ad serving platforms use routing logic to direct bid requests to the campaigns best matched for each impression opportunity. The AI-specific use of the term typically refers to learned routing within neural network architectures rather than infrastructure-level routing.
Mixture-of-experts models with dynamic routing are used in many of the AI tools a working ad agency uses for content generation, research, and analysis. Understanding at a conceptual level how these models allocate computation helps agencies understand why these tools can be so capable while remaining commercially viable to run, and what that capability profile implies for agency applications.
Specialization has implications for output consistency. When a model routes different inputs to different experts, the handling of similar inputs is not guaranteed to be consistent if they cross an expert boundary. Some output variability in large language models traces to routing decisions at the edge of expert territories. This is useful context when evaluating why a model handles very similar prompts differently in ways that seem arbitrary.
It informs vendor conversations about model size claims. A vendor claiming their model has one trillion parameters is making a different statement than it appears if the model is a mixture-of-experts architecture with active parameter counts an order of magnitude lower. Understanding that total parameters and active parameters per input are different numbers helps agencies interpret vendor benchmarks more critically.
Agent architectures use routing logic. AI agent systems that direct tasks to different specialized tools or models based on the nature of each request are implementing a form of dynamic routing at the application level. The same principle, specialized subcomponents selected by a gating mechanism, applies whether the routing is learned weights in a neural network or conditional logic in an agent workflow.
An agency is selecting a large language model for a multi-purpose workflow that includes long-form content generation, structured data extraction from research documents, and code generation for campaign automation scripts. Initial testing reveals that different model architectures perform differently across these three task types: the model that scores highest on content generation is not the highest scorer on structured extraction or code generation. A review of the model architecture documentation reveals that one candidate uses a mixture-of-experts approach with domain-specialized experts, which accounts for its consistent high performance across all three categories: different expert subnetworks handle each task type. The agency selects it for the multi-purpose workflow and notes the routing architecture as a factor in vendor selection documentation.
The generative AI foundations module of the workshop covers how today’s models work, including the architectural differences that determine when a model is the right tool for a specific agency task.