Breaking Babel: A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation
Haibo Jin, Xinjie Li, Najmeh Sadoughi, Yang Liu, Yibo Wang, Zhu Liu, Yuzong Liu
University of Illinois Urbana-Champaign · Amazon
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 35 · arXiv 2609.38660
The abstract
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. We propose SMART, a Self-evolving Multi-Agent system for long-foRm subtitle Translation. During test-time training, SMART builds persistent series-level memory and translates a subset of sentences through a dynamic router and Mixture-of-Agents layer with tools for terminology verification, subtitle constraint validation, and contextual retrieval. A judge-refiner loop scores candidates and uses textual critiques to update agent prompts and routing policies without retraining the underlying LLMs. During test-time inference, the evolved configuration translates the remaining series. We also introduce Subtitle Arena, covering 14 genres, 2--198 episodes per series, production years 1959--2023, and 15 target locales, together with SubMQM, a subtitle-adapted MQM framework with seven dimensions and 19 error categories. SMART achieves the best overall MQM score in all 15 Subtitle Arena directions, reducing average penalty by 6.9% over the strongest competing agent system. On the public MuSC benchmark, SMART obtains the best model result across all four language pairs and also achieves the best human-evaluation result, with an overall score of 4.50/5.
The Brief
2:28A short podcast summary of the paper. Written and voiced by AI.
Transcript
Breaking Babel. A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation. By Haibo Jin and six others, from the University of Illinois Urbana-Champaign and Amazon. Published on the twenty-ninth of September. No code repository is linked to it.
Translating the subtitles of a whole series is not a sentence-by-sentence job. It needs reasoning over discourse and cultural context that can stretch across episodes or across the entire run of a show, and it has to keep terminology and style consistent the whole way. Existing methods built on a single large language model mostly work at the level of the sentence. Multi-agent systems go further, but they often follow static workflows that do not adapt to how complex a scene is or to the context of the production.
SMART, a Self-evolving Multi-Agent system for long-form subtitle Translation, lets the system adapt itself to the series in front of it. It splits the work into two phases. In a test-time training phase it learns from part of the series. In a test-time inference phase the configuration it evolved translates the rest.
During the training phase, SMART builds a persistent memory at the level of the series, and translates a subset of sentences through a dynamic router and a Mixture-of-Agents layer. That layer has tools for verifying terminology, validating subtitle constraints, and retrieving context. A judge-refiner loop scores the candidate translations and turns textual critiques into updates to the agent prompts and routing policies. The underlying language models are never retrained. What evolves is the prompts and the routing.
Alongside the system come two evaluation resources. Subtitle Arena covers fourteen genres, from two to one hundred and ninety-eight episodes per series, productions spanning more than six decades, and fifteen target locales. SubMQM adapts the MQM framework to subtitles, with seven dimensions and nineteen error categories.
On Subtitle Arena, SMART gets the best overall MQM score in all fifteen directions, cutting the average penalty by six point nine percent relative to the strongest competing agent system. On the public MuSC benchmark, it gives the best model result on all four language pairs, and the best human-evaluation result too, with an overall score of four point five out of five.
Where SMART departs from a static pipeline is in when the learning happens. Nothing about the models changes. Instead, a slice of each series is spent teaching the agent system how to handle that particular series, and only then does it translate the remainder with what it has written down for itself.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A series is not a pile of sentences. Translating it like one shows.
You are translating the subtitles of a whole TV series. A line's meaning can hinge on a scene three episodes back, names and terms must stay the same throughout, and every line must fit on screen.
The road ahead
- 1 · The problem ~6 min Translating a whole series means remembering context across episodes, keeping names fixed and fitting the screen, and today's translators treat every line alike.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Let the system remember the series, choose helpers per line, and rewrite its own prompts and routing from a judge's critiques, without retraining any model.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Translate part of the series through routed agents, judge the drafts, rewrite prompts and routing from critiques, then translate the rest with the frozen setup.
- 6 · The experiment ~6 min You’ll guess: Against the strongest competing agent system on Subtitle Arena, by how much did SMART reduce the average MQM penalty?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.