A Dual-Track Framework for Template-Constrained LaTeX Conversion

Converting structured Markdown drafts into venue-specific LaTeX templates is a persistent challenge in document engineering. Existing solutions typically fall into two categories: rule-based converters (like Pandoc), which lack the contextual reasoning to handle complex template constraints, or end-to-end LLM generation, which is prone to structural hallucinations, semantic drift, and high computational costs. This paper addresses the need for a system that simultaneously ensures structural fidelity, semantic preservation, and robust compilation.

The authors propose a Dual-Track Framework that separates template comprehension from document execution. The system operates through two primary tracks:

• Offline Track: An LLM analyzes target LaTeX template files to extract formatting constraints and package requirements, serializing them into a reusable manifest. This eliminates the need for repeated template reasoning during runtime.
• Online Track: A hybrid execution pipeline processes the Markdown input. It segments the document into functional regions—deterministic, probabilistic, and neuro-symbolic—and routes each to the most appropriate engine. LLM invocation is strictly limited to reasoning-intensive components like bibliographic references and complex layouts, while deterministic syntax is handled by rule-based engines.

Evaluated on 56 published research papers across 7 academic templates, the framework demonstrates significant improvements over existing baselines. It achieved a 94.6% compilation success rate, compared to 82.1% for rule-based methods and 62.5% for monolithic LLM pipelines. Furthermore, by bounding neural generation to targeted regions, the framework achieved a 5.1x speedup and substantially reduced token consumption, providing a more stable and cost-effective solution for production environments.