Large Language Models (LLMs) have shown impressive capabilities in reasoning, but they continue to fail in complex, multi-step planning tasks. These failures frequently manifest as constraint violations, inconsistent state tracking, and brittle solutions that break under minor changes.
We argue that many of these issues arise not from a lack of reasoning capability itself, but from the absence of an explicit problem representation. Current prompting paradigms allow reasoning to proceed over an implicit and unstable internal model of the task, which is highly prone to drift and contradiction as the reasoning length increases.
In contrast to these traditional approaches, human scientific and engineering reasoning is fundamentally model-based: we define relevant entities and governing laws before drawing inferences. To emulate this, we propose Model-First Reasoning (MFR), a two-phase paradigm that explicitly separates problem representation from problem solving. In the first phase, the model is required to construct a structured model of the problem space, identifying all relevant entities, state variables, and constraints. Crucially, the model is instructed not to generate any solution steps during this initial phase.
Experiments across diverse, constraint-driven domains such as medical scheduling and resource allocation show that MFR substantially reduces errors compared to standard methods. By forcing the model to articulate its understanding of a problem before solving it, MFR makes the representation inspectable and correctable.
Only after this modelling phase is complete does the model proceed to reasoning and planning. All subsequent steps are then conducted strictly with respect to the constructed model: actions must respect stated preconditions, and transitions must be consistent with defined effects. This separation introduces a representational scaffold that constrains the model and reduces its reliance on unstable latent state tracking. Because the model is externalised and made explicit, any logical violations become visible and diagnosable, allowing for easier verification by both humans and automated systems.
Experiments across diverse, constraint-driven domains such as medical scheduling and resource allocation show that MFR substantially reduces errors compared to standard methods. By forcing the model to articulate its understanding of a problem before solving it, MFR makes the representation inspectable and correctable. This approach reframes hallucinations not just as false statements, but as a symptom of reasoning performed without a clearly defined model of the problem space. MFR does not require architectural changes or additional training, making it a foundational and immediately applicable component for building more reliable, agentic AI systems.
The team at Academii are always happy to discuss all your training and education needs, help your organisation attract and train new talent, and build a resilient workforce. Please drop us a line here to know more.
