SMART turns full-season subtitle translation into a stateful long-form task and posts the lowest SubMQM penalty across 15 directions

SMART models long-form subtitle translation as a stateful, adaptive process: test-time training evolves agent prompts, routing policy, and series-level memory, while test-time inference freezes the configuration but lets memory keep accumulating. Unlike multi-agent translation systems with fixed workflows, SMART evolves at test time; unlike general self-evolving agents, it targets long-form translation, improving future sentences while preserving terminology, style, and prior translation competence; unlike conventional prompt optimization, it treats prompts and agent behaviors as persistent system state updated through translation-specific feedback. The paper specifies the two-stage procedure, the system state (translator pool, prompts, routing policy, series memory), and a five-phase test-time training loop (TRANSLATE, EVALUATE, PROMPT OPTIMIZATION, STRUCTURE OPTIMIZATION, SAVE CONFIG), noting that prompt and routing updates are constrained, applied once per training epoch, and leave the underlying LLM parameters unchanged.

Ablations show series-level memory is the largest contributor, raising the average Overall penalty by about 0.78 when removed, followed by MoA, contextual retrieval and the idiom bank, sliding-window consistency, self-evolution, and dynamic routing. The result separates the role of long-range context, knowledge retrieval, and multiple specialist hypotheses from instruction content alone: replacing MoA with a Combined Prompt raises the penalty by about 0.36, indicating the gain comes from independently generated specialist hypotheses rather than the instructions themselves. Ablations run across six representative directions (enzh, ende, enko, enit, enes, enfr) with per-setting penalty increments, and each variant changes only the indicated component while retaining the remaining SMART configuration and evaluation protocol.

SMART obtains the lowest Overall SubMQM penalty in all 15 English-to-locale directions and all 15 reverse directions on Subtitle Arena, cutting the average penalty by 6.9% relative to the strongest agent baseline, TransAgent, with improvements spread across semantic and subtitle-specific dimensions rather than one error type. Relative to TransAgent, the mean Accuracy penalty and mean Technical penalty both fall, while Terminology, Fluency, Linguistic Conventions, Locale Conventions, and Audience Appropriateness also improve; the reverse locale-to-English evaluation shows the same overall trend, indicating the gains are not specific to generating diverse target languages. Results use SubMQM, a subtitle-adapted MQM protocol with seven dimensions and 19 error types, with the same evaluator and rubric for all systems and lower penalties being better; the paper also points to complete fine-grained results for all 30 directions in the appendix.

On the public MuSC benchmark, SMART achieves the best model result on Accuracy, Naturalness, and Vividness across all four language pairs, and in a blind study with 20 annotators it ranks first on all four criteria with an overall score of 4.50/5. Gains are especially clear in Vividness, indicating that high-quality subtitles require contextual and stylistic adaptation over long-form narrative content, not only semantic fidelity; in the human study SMART performs best on Consistency. MuSC results are compared against strong reasoning models and the fine-tuned ALPO baseline; the human study uses aligned subtitle passages with corresponding video, five anonymized candidates in randomized order, and best-to-worst ranking on Fidelity, Consistency, Language, and Subtitle, with the paper noting the study's limited scale and coarse criteria and retaining SubMQM for fine-grained error analysis.

The result targets long-form, continuously narrative subtitle translation, especially TV-series localization that must maintain terminology, character references, and style across episodes; the paper states SMART is most suitable where semantic accuracy, contextual consistency, and subtitle quality matter more than minimizing inference-time computation, and it demonstrates applicability with weaker translation backbones, an alternative evaluator, added video and audio multimodal tools, and 2025–2026 series. Subtitle Arena and SubMQM provide a series-level, error-typed evaluation basis that can support further comparison of discourse preservation, terminology consistency, and subtitle display constraints.

The loaded text is the full paper, but some appendix tables are truncated in the evidence bundle, for example the enfr row of the multimodal expansion table, so per-direction multimodal details can only be read through the main-text summary (average SubMQM penalty falling from 0.63 to 0.57 across six representative directions and outperforming ViDove and Hermes). In addition, the human study uses 20 annotators and four coarse criteria, which the paper itself frames as validating viewer-facing quality rather than replacing fine-grained diagnostics; the temporal generalization evaluation uses 200 series released in 2025–2026 and reports comparisons against TransAgent across six directions. Readers interested in specific error types or a particular locale would still need the complete fine-grained appendix tables.