HULAT2 at MER-TRANS 2026: Governed Multi-Agent Simplification for Spanish Easy-to-Read Generation

This paper details the participation of HULAT2-UC3M in the Spanish track of MER-TRANS 2026, a shared task on multilingual Easy-to-Read translation. The team submitted three fully automatic runs comparing a multi-agent workflow against a linear baseline to evaluate simplification strategies.

• RUN1 and RUN2 utilized a LangGraph-based multi-agent workflow combining Gemini 2.5 Flash and RigoChat-7B-v2 with Event-Condition-Action routing.
• RUN2 added an additional lexical-support layer based on a glossary and lexical resources to the base workflow.
• RUN3 served as a baseline using a RigoChat-based generate-evaluate-regenerate approach with prompt engineering and LoRA-based adaptation.
• Official SARI scores ranked RUN1 first with 44.0543 points, followed by RUN2 at 43.1049 and RUN3 at 38.5136.

The results indicate that signal-guided multi-agent routing outperformed the linear regeneration baseline in this task setting. However, adding lexical support did not automatically improve reference-based scores, suggesting a need for further analysis of readability and factual consistency.