AI
Breaking Babel: A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 10:46:30 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 05:18:24 UTC.
Briefings
SMART is a self-evolving multi-agent system designed for long-form subtitle translation, achieving the best MQM score in the Subtitle Arena across 15 genres and 15 target locales. It uses a judge-refiner loop to update agent prompts dynamically during inference, reducing average penalties by 6.9% compared to competing systems. SMART also achieved the best human-evaluation result on the MuSC benchmark with a score of 4.50/5.
Key points
- SMART employs a dynamic router and Mixture-of-Agents layer to translate a subset of sentences during test-time training, while the evolved configuration handles the remaining series during inference.
- The system maintains persistent series-level memory and utilizes a judge-refiner loop to update routing policies without retraining the underlying LLMs, achieving a 6.9% reduction in average penalties over competitors.
Context and limitations
The study covers 14 genres, 2-198 episodes per series, production years 1959-2023, and 15 target locales, with limited coverage on specific episode complexities and production contexts.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 10:46:30 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeie4iz2kbdzn6sq44lk65retina3rn6sjxcbogd6f75z33zbjfk7oq