Model-First Reasoning LLM Agents: Reducing Hallucinations through Explicit Problem Modeling

Abstract

Large language models often struggle with complex multi-step planning tasks, showing high rates of constraint violations and inconsistent solutions. The authors propose Model-First Reasoning (MFR), a two-phase paradigm where the LLM first constructs an explicit problem representation — entities, state variables, actions, constraints — before generating any solution. The core claim is that many LLM planning failures stem from representational deficiencies rather than reasoning limitations.

Key Contributions

• MFR paradigm — separates problem modeling from solution generation
• Explicit state tracking instead of implicit chain-of-thought state management
• Reduced constraint violations across five planning domains
• Diagnosis of planning failures as representational, not reasoning-bound

Methodology

Two phases:

1. Modeling phase — LLM constructs a structured representation defining:
   • Entities in the problem
   • State variables
   • Actions available
   • Constraints that must hold
2. Solution phase — LLM generates a plan conditioned on the explicit model built in phase 1

Ablation studies validate the criticality of the modeling phase — removing or compressing it collapses performance.

Results

MFR outperforms Chain-of-Thought and ReAct baselines across multiple domains:

• Medical scheduling
• Route planning
• Resource allocation
• Logic puzzles
• Procedural synthesis

Specific reduction in constraint violations relative to CoT/ReAct reported across all five domains.

Limitations

Abstract does not detail failure modes quantitatively. Open questions: how does MFR compare on open-ended tasks without crisp constraint formulations? Does the explicit-modeling step scale to real-world agentic environments where state is partially observable?

Source: Model-First Reasoning LLM Agents — Annu Rana, Gaurav Kumar, December 2025