Breaking Babel: A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation

The authors propose SMART, a Self-evolving Multi-Agent system for long-form subtitle translation that builds persistent series-level memory during test-time training. SMART coordinates a dynamic router and a Mixture-of-Agents layer with tools for terminology verification, constraint validation, and contextual retrieval, using a judge-refiner loop to update agent prompts and routing policies without retraining the underlying LLMs. Evaluated across 15 target locales in Subtitle Arena, SMART achieves the best overall MQM score, reducing average penalties by 6.9% over the strongest competing agent system, and secures the top human-evaluation score of 4.50/5 on the public MuSC benchmark.

SMART achieves the best overall MQM score in all 15 Subtitle Arena directions.

It reduces average penalty by 6.9% over the strongest competing agent system.

SMART secures an overall human-evaluation score of 4.50/5 on the public MuSC benchmark.