Breaking Babel: A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation
SMART adapts its agents at test time before translating the rest of a subtitle series.
The paper says the system builds persistent series-level memory, routes harder subtitle lines through specialized agents, and uses tools for terminology, constraints, and retrieval. A judge-refiner loop critiques candidate translations, then updates prompts and routing policies without retraining the base LLMs. The authors also introduce Subtitle Arena and SubMQM to evaluate long-form subtitle translation across 15 locales. SMART reports the best MQM score in all 15 directions, with a 6.9% lower average penalty than the strongest competing agent system. HF Daily Papers' note
The paper says the system builds persistent series-level memory, routes harder subtitle lines through specialized agents, and uses tools for terminology, constraints, and retrieval. A judge-refiner loop critiques candidate translations, then updates prompts and routing policies without retraining the base LLMs. The authors also introduce Subtitle Arena and SubMQM to evaluate long-form subtitle translation across 15 locales. SMART reports the best MQM score in all 15 directions, with a 6.9% lower average penalty than the strongest competing agent system. HF Daily Papers' note
score 4