Opening — Why this matters now
As LLM agents graduate from clever chatbots to decision‑making systems, their failures are becoming less amusing and more expensive. We are no longer talking about wrong trivia answers; we are talking about broken schedules, invalid plans, unsafe workflows, and agents confidently violating constraints they were never told—explicitly—not to break.
The default response has been predictable: better prompts, longer Chain‑of‑Thoughts, more tool calls, tighter loops. And yet, the same failures persist. The uncomfortable implication is that we may be fixing the wrong layer.
This paper argues exactly that—and it does so with refreshing clarity.
Background — Reasoning without a model is just improvisation
Modern LLM agent paradigms like Chain‑of‑Thought (CoT) and ReAct have dramatically improved multi‑step reasoning. They encourage models to “think aloud,” interleave actions with observations, and simulate deliberation. What they do not require is something far more basic: an explicit statement of what the problem actually is.
Entities appear when convenient. Constraints are remembered—until they aren’t. State exists as a diffuse narrative memory scattered across tokens. Locally, everything sounds coherent. Globally, things quietly break.
This is not how humans reason when stakes matter. Scientists define variables before drawing conclusions. Engineers formalize systems before optimizing them. Classical AI planners refused to even start without a domain model.
LLM agents, by contrast, are often asked to plan first and understand later.
Analysis — Model‑First Reasoning (MFR)
The core proposal of the paper is disarmingly simple: separate modeling from reasoning.
Model‑First Reasoning (MFR) enforces a two‑phase process:
-
Model construction — The LLM must explicitly define:
- Entities
- State variables
- Actions (with preconditions and effects)
- Constraints
No solutions. No plans. Just structure.
-
Reasoning over the model — Only after the model is fixed does the agent generate a plan, strictly constrained by what it has already declared.
This is not symbolic planning. There is no PDDL parser lurking in the background. The model can be expressed in plain language. What matters is that it is externalized, inspectable, and stable.
In effect, MFR forces the LLM to commit to a world before acting within it.
Findings — What changes when structure comes first
Across a range of constraint‑heavy tasks—medical scheduling, routing with temporal dependencies, resource allocation—the differences are stark.
| Reasoning Strategy | Constraint Violations | Implicit Assumptions | Structural Clarity |
|---|---|---|---|
| Chain‑of‑Thought | Medium | Frequent | Low |
| ReAct | Medium–Low | Occasional | Medium |
| Model‑First | Low | Rare | High |
Three patterns stand out:
- Constraint grounding improves immediately once constraints are named upfront.
- Assumption creep collapses when actions and states must be declared explicitly.
- Plans become debuggable, not just plausible‑sounding.
Perhaps most telling: the reasoning itself does not become more sophisticated. It becomes more honest.
Implications — Hallucination is a modeling failure
The paper reframes hallucination in a way that should unsettle anyone building agentic systems: many failures are not inference errors, but representational ones.
When an agent invents a step, skips a dependency, or violates a rule, it is often because that rule was never stabilized as part of its internal—or external—model. Asking for better reasoning on top of an unstable representation is like demanding accurate navigation without a map.
MFR offers a form of soft symbolic grounding: enough structure to constrain behavior, without sacrificing the flexibility that makes LLMs useful.
For businesses deploying agents in regulated, safety‑critical, or high‑cost environments, this matters. Interpretability, auditability, and reproducibility are not “nice‑to‑haves.” They are prerequisites.
Conclusion — Think less, model more
Model‑First Reasoning does not require new architectures, retraining, or external solvers. It requires something more uncomfortable: slowing down and insisting on clarity before action.
The lesson is almost embarrassingly old‑fashioned. Reasoning does not create structure; it operates on it. If the structure is implicit, unstable, or incomplete, no amount of eloquent thinking will save you.
In a field obsessed with making models think harder, this paper reminds us to make them understand the problem first.
Cognaptus: Automate the Present, Incubate the Future.