MAUL: The Hypothesis
The hypothesis is simple: specialization beats generalization. A single model doing everything — math, code, search, open-ended reasoning — has to compromise on each. It cannot be an expert at all of them simultaneously. MAUL is built on the premise that you should not ask it to be.
The problem with generalist models
Language models fail at math not because they lack intelligence — they fail because math requires determinism, and language models are probabilistic. When you ask a model what 847 × 63 is, it pattern-matches to sequences it has seen before. It does not compute. It guesses, and sometimes the guess is right.
The same failure mode appears in code execution, real-time information retrieval, and anything requiring external state. A single model handling all of these is forced into compromises that none of these domains can afford.
The routing approach
MAUL reads what you asked, determines which path gets you the correct answer, routes to the right specialist, and returns one reply. The coordination is invisible to the user.
The three primary paths:
- Calculator path. Math problems route through a deterministic calculator (SymPy). The model does not compute — it extracts the expression and hands it to something that can. Hallucination on arithmetic drops to near zero.
- Search path. Questions requiring external context — current events, facts the model cannot know — route through web search and Wikipedia retrieval. The model synthesises the retrieved content into a response.
- Direct path. Reasoning, creative, and conversational requests go directly to the appropriate specialist agent — Coder, Logic, Creative, Chat — each with prompts tuned to their domain.
Semantic understanding before routing
MAUL-2 introduced a semantic understanding phase that runs before any routing decision. The system interprets intent — not just keywords — before determining the path. A question phrased as “what is the square root of 144” and one phrased as “144 under the radical sign” should route identically. They do.
Results
MAUL-2 paired with Hydrogen-2: 95.90% on GSM8K, 97.56% on HumanEval. These are the scores after routing math through the calculator and code through a sandboxed executor — not after asking a model to do its best.
What comes next
MAUL-3 adds a ReACT engine — the model reasons step by step before acting, and can loop back when an intermediate result is wrong. It also adds session memory so context persists across turns. The routing hypothesis holds across versions; the architecture around it becomes more capable.
The core belief behind MAUL has not changed: the right tool for the right job, coordinated by a system smart enough to know which tool that is.