Proposes SMART, a self-evolving multi-agent framework that dynamically adapts workflows for long-form subtitle translation based on scene and production context.
SMART (Self-evolving Multi-Agent system for long-foRm subtitle Translation) is a framework introduced in the September 2026 arXiv paper Breaking Babel to address the limitations of static, sentence-level machine translation in long-form subtitles. It operates via a test-time training stage where it builds a persistent series-level memory and uses a judge-refiner loop to update agent prompts and routing policies without retraining the underlying LLMs.
The system employs a dynamic graph router and a Mixture-of-Agents layer to select specialized translators and tools for terminology verification, constraint checking, and contextual retrieval. This adaptive approach allows SMART to maintain consistent terminology, style, and cultural context across entire series, outperforming existing single-LLM and static multi-agent methods.
Evaluated on the Subtitle Arena benchmark across 15 target locales, SMART achieved the best overall SubMQM scores, reducing the average penalty by 6.9% compared to the strongest competing agent system. It also demonstrated significant improvements in vividness and reduced errors in accuracy, fluency, and locale conventions, proving effective for complex, context-dependent translation tasks.
Breaking Babel presents SMART, a self-evolving multi-agent system for translating long-form subtitles, framing the task as a context-sensitive workflow problem rather than a single large language model generation. Long-form media introduces persistent constraints—character voice, episode-level continuity, scene-specific register, timing, terminology, and production conventions—that are difficult to capture with static prompt templates or isolated translation calls. SMART addresses this by decomposing the pipeline into cooperating agents that can extract scene context, maintain translation state, draft and revise subtitles, and enforce quality constraints, while dynamically selecting or reconfiguring workflow steps according to the input.
The paper’s central contribution is the self-evolving orchestration mechanism: instead of using a fixed agent graph, SMART adapts its workflow over time based on feedback, accumulated context, and observed failure modes. This may involve updating prompts, revising agent responsibilities, strengthening terminology or style memories, and prioritizing different validation steps for different scenes or production settings. The key insight is that high-quality long-form subtitle translation benefits from a system that learns not only from individual examples but from the structure of the translation process itself, allowing it to improve consistency, reduce rework, and better handle heterogeneous narrative content.
The work matters because it points toward a more practical architecture for AI-assisted localization in complex, long-context domains. Rather than relying on monolithic models or brittle hand-designed pipelines, a self-evolving multi-agent system can better approximate the iterative, context-aware behavior of professional translation teams. For technically literate readers, the significance lies in its treatment of LLM agents as an adaptive control system: the paper suggests that future media translation tools should be evaluated not only on isolated sentence quality, but on their ability to maintain coherence, adapt to production constraints, and improve workflow decisions across extended content.