1. What is a compound model?
A compound model is a model made up of several other models. Most compound models simply collect a response from every participating model and merge them into one. Leviathan does more than that.
2. Leviathan
Leviathan is a single system built on several frontier models: Anthropic's Claude Opus 4.8, OpenAI's GPT 5.5, Google's Gemini 3.1 Pro, and DeepSeek V4 Pro. Instead of merging their responses, it uses a dedicated mechanism to improve their reasoning.
3. General-to-specific reasoning
We call that mechanism general-to-specific reasoning (GSR). A model first explains every concept related to the query, moving from the broadest concepts to the most specific. This grounds its answer more firmly in knowledge the model already has.
4. Debate as a path to truth
John Stuart Mill argued, controversially, that debate can be a source of truth. The idea holds up well for AI, because models lack the human drive to win at all costs, so they can debate properly. Each model can correct flaws in another's reasoning and bring in information the others missed, which produces a more thorough understanding.