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

2026-07-02 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Natural Language Processing · Depth: Expert, quick

Summary

HULAT2-UC3M participated in the Spanish track of MER-TRANS 2026, a shared task focused on multilingual Easy-to-Read translation, submitting three fully automatic Spanish runs. RUN1 and RUN2 utilized a LangGraph-based multi-agent workflow, integrating Gemini 2.5 Flash and RigoChat-7B-v2, with parallel generation, internal quality signals, Event-Condition-Action routing, controlled editing, and traceable decisions. RUN2 added a lexical-support layer, while RUN3 served as a RigoChat-based generate-evaluate-regenerate baseline with prompt engineering and LoRA adaptation. Official SARI scores reported RUN1 as the best HULAT2 entry with 44.0543 points, followed by RUN2 at 43.1049, and RUN3 at 38.5136. These results indicate that signal-guided multi-agent routing surpassed the linear regeneration baseline, and that adding lexical support did not automatically enhance reference-based scores in this specific task.

Key takeaway

For NLP Engineers developing Easy-to-Read text generation systems, this research suggests prioritizing signal-guided multi-agent architectures over simpler generate-evaluate-regenerate baselines. You should explore LangGraph-based workflows with diverse LLMs and internal quality signals for routing decisions. Be cautious with adding lexical support, as it may not automatically boost reference-based metrics like SARI, requiring deeper analysis for user-oriented adequacy.

Key insights

Signal-guided multi-agent routing improved Easy-to-Read generation over a linear baseline for Spanish.

Principles

• Multi-agent routing can enhance text simplification.
• Lexical support doesn't guarantee score improvement.
• Internal signals guide complex generation workflows.

Method

The method involved a LangGraph-based multi-agent workflow combining Gemini 2.5 Flash and RigoChat-7B-v2, parallel generation, internal quality signals, Event-Condition-Action routing, controlled editing, and traceable decisions.

In practice

• Implement LangGraph for multi-agent LLM workflows.
• Integrate internal quality signals for routing.
• Test lexical resources' impact on specific metrics.

Topics

• Easy-to-Read Generation
• Multi-Agent Systems
• LangGraph
• Text Simplification
• Large Language Models
• MER-TRANS 2026

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Building an Intelligent Natural Language to SQL Pipeline Using LangGraph and Multi-Agent… — A multi-agent LangGraph architecture enhances NL-to-SQL accuracy and safety by decomposing tasks and integrating human oversight.
• When Cultures Meet: Multicultural Text-to-Image Generation — Multi-agent LLM frameworks improve multicultural image generation quality and alignment by creating richer, contextually nuanced captions.
• InterleaveThinker: Reinforcing Agentic Interleaved Generation — InterleaveThinker uses a multi-agent pipeline to enable existing image generators to perform complex interleaved text-image generation.
• The Sequence Knowledge #866: Three Text Diffusion Models You Need To Know About — Text diffusion models iteratively refine noisy or masked text, challenging traditional sequential language generation.
• Google's PaperBanana uses five AI agents to auto-generate scientific diagrams — A multi-agent AI system automates scientific diagram generation, improving aesthetics but still facing content accuracy challenges.
• How Generation Architecture Shapes Code Complexity in Multi-Agent LLM Systems: A Paired Study on HumanEval — Multi-agent LLM code generation architectures significantly impact code complexity without necessarily improving functional correctness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.