Research
SERV Reasoning:
Made for Production Agents
The paper behind SERV Reasoning.
Published at arXiv:2512.15959:
The thesis
English is great for communication.
It is bad for computation.
Frontier models reason in tokens. A token is a fragment of language. Language was built for ambiguity, indirection, and context — none of which compose.
A reasoning system that lives at agent scale cannot afford to spend every step in natural language. The graph has to be in the graph, the schema has to be in the schema, the routing has to know which model is fit-for-purpose.
This is why the cost curve refuses to bend with prompting alone.
Three contributions.
One reasoning architecture.
Level 1
Bounded reasoning graphs
A formal specification for decomposing tasks into explicitly-typed reasoning steps with dependency contracts.
Level 2
Schema-forced execution
A constraint mechanism that eliminates parse failures and bounds inference cost without re-prompting.
Level 3
Intelligent model routing
A learned policy for assigning reasoning sub-tasks to the model class fit-for-purpose.
Authors.
The people behind the work.
Publications.
Papers, posts, and citations from the field.
🍔 Our Publications