Breaking Babel: A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation
SMART, a self-evolving multi-agent translator, tops Subtitle Arena and MuSC by adapting prompts and routing at test time.
SMART is a self-evolving multi-agent system for long-form subtitle translation that builds series-level memory and routes sentences through a Mixture-of-Agents layer with terminology, constraint, and retrieval tools. A judge-refiner loop updates agent prompts and routing policies at test time without retraining the underlying LLMs. On Subtitle Arena, covering 14 genres, 15 locales, and series from 1959 to 2023, SMART records the best MQM score in all 15 directions and cuts average penalty by 6.9% versus the strongest competing agent system. On public MuSC it leads all four language pairs and scores 4.50/5 in human evaluation.
- SMART updates prompts and routing at test time without retraining LLMs.
- Best MQM score in all 15 Subtitle Arena translation directions.
- Average penalty falls 6.9% versus the strongest competing agent system.
- Best MuSC result on four language pairs, with human score 4.50/5.
From the original article
1 min read at huggingface.co
Long-form subtitle translation requires reasoning over discourse and cultural context spanning episodes or entire series, while maintaining consistent terminology and style. Existing single-LLM methods are largely sentence-level, and multi-agent systems often use static workflows that do not adapt to scene complexity or production context.
By Haibo Jin, Xinjie Li, Najmeh Sadoughi · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.