Translation Agent in Modular LLM Systems

• Translation Agent is a specialized module in LLM systems that handles bounded subtasks like drafting, refinement, and evaluation.
• Systems with translation agents break down translation into stages such as retrieval, grounding, and iterative repair to enhance quality and consistency.
• Applications span natural language, code modernization, simultaneous speech, and legal translations, offering measurable improvements in semantic fidelity and structural correctness.

Searching arXiv for the key papers to ground the article in current literature. A translation agent is a modular, role-specialized component in an LLM-based translation system that performs a bounded subtask within a larger workflow, rather than attempting end-to-end translation in a single pass. In current research, the term covers both natural-language and code-translation settings, including literary translation, document-level machine translation, simultaneous speech translation, multimodal subtitling, legal translation, LaTeX translation, and repository-scale code migration. Across these settings, translation agents are typically embedded in multi-agent systems that decompose translation into drafting, retrieval, grounding, validation, critique, revision, or orchestration stages, with the stated goal of improving semantic fidelity, structural correctness, consistency, and domain conformity under real deployment constraints (Li et al., 10 Jun 2025, Moti et al., 14 Mar 2026, Wang et al., 2024, Yuan et al., 2024).

1. Concept and scope

In the contemporary literature, a translation agent is not defined by a single model architecture but by functional specialization. In TACTIC, an agent handles one cognitive subtask such as drafting, refinement, evaluation, scoring, context reasoning, or knowledge gathering (Li et al., 10 Jun 2025). In LegacyTranslate, the “Initial Translation Agent” generates a first Java draft from PL/SQL using retrieved exemplars, while other agents ground the draft in internal APIs and repair it through compiler feedback (Moti et al., 14 Mar 2026). In DelTA and Loong, the translation agent is document-level and memory-centric: it translates incrementally while consulting summaries, entity records, exemplars, and prior sentence pairs to maintain coherence across long documents (Wang et al., 2024, Wang et al., 28 May 2026). In CLASI and Agent-SiMT, the translation agent operates under streaming constraints, producing output from partial input while separate components control read/write timing or chunk boundaries (Cheng et al., 2024, Guo et al., 2024).

This breadth of usage has made “translation agent” a cross-domain term. In code migration, the agent often translates source units under compilation, test, and framework constraints rather than lexical equivalence alone (Guan et al., 25 Aug 2025, Ibrahimzada et al., 8 Apr 2026). In natural-language translation, the same term often denotes agents that reason over context, style, terminology, or evaluation criteria (Chen et al., 2024, Wang et al., 20 May 2025). A plausible implication is that the field treats translation less as a single decoding act and more as a controlled process in which multiple specialized agents expose intermediate structure that monolithic prompting tends to hide.

2. Core architectures and coordination patterns

Most translation-agent systems use explicit role decomposition. Some systems are sequential pipelines, some are iterative repair loops, and some introduce routing or adversarial interaction. LegacyTranslate uses three agents: Initial Translation, API Grounding, and Refinement (Moti et al., 14 Mar 2026). RepoTransAgent uses a RAG Agent, a Context Agent, and a Refine Agent (Guan et al., 25 Aug 2025). TRANSAGENT uses four agents—Initial Code Translator, Syntax Error Fixer, Code Aligner, and Semantic Error Fixer—under a controller that alternates between validation and localized repair (Yuan et al., 2024). TACTIC uses six agents aligned with Cognitive Translation Studies: DraftAgent, RefinementAgent, EvaluationAgent, ScoreAgent, ContextAgent, and ResearchAgent (Li et al., 10 Jun 2025). TransLaw uses Translator, Annotator, and Proofreader (Xuan et al., 1 Jul 2025). MAATS replaces a single reviewer with multiple MQM-specialized evaluators followed by an overview editor (Wang et al., 20 May 2025).

System	Domain	Specialized agents
LegacyTranslate	PL/SQL→Java modernization	Initial Translation, API Grounding, Refinement
RepoTransAgent	Repository-aware code translation	RAG, Context, Refine
TRANSAGENT	Code translation	Initial Code Translator, Syntax Error Fixer, Code Aligner, Semantic Error Fixer
TACTIC	Machine translation	Draft, Refinement, Evaluation, Score, Context, Research
TransLaw	Legal translation	Translator, Annotator, Proofreader
MAATS	Machine translation	Translator, MQM evaluators, Editor

These decompositions differ in what they treat as the hard part of translation. Some systems assume generation is easy but timing or context selection is difficult, as in simultaneous translation (Guo et al., 2024). Others assume initial generation is easy but environment integration is difficult, as in enterprise code modernization (Moti et al., 14 Mar 2026). Others treat evaluation itself as a first-class agentic task, as in MAATS and DITING’s AgentEval, where multiple evaluators simulate deliberation rather than relying on a single scalar metric (Wang et al., 20 May 2025, Zhang et al., 10 Oct 2025).

An important architectural distinction concerns control. Some frameworks use fixed workflows, such as Translator→Annotator→Proofreader in TransLaw (Xuan et al., 1 Jul 2025). Others use dynamic routing. SAMAS computes an 81-dimensional Stylistic Feature Spectrum and deterministically chooses a “Faulkner-esque” or “Hemingway-esque” workflow based on thresholds $H > 0.85$ and $E_{\text{low}} > 0.6$ (Wu et al., 23 Feb 2026). Loong performs stepwise observe-and-act context selection before translating each segment (Wang et al., 28 May 2026). ACToR uses adversarial collaboration: a translator agent repairs Rust code while a discriminator agent searches for new failing tests (Li et al., 4 Oct 2025).

3. Context, memory, and grounding mechanisms

A defining property of translation agents is explicit context management. Direct prompting is repeatedly described as inadequate when translation depends on repository structure, document discourse, internal APIs, or multimodal evidence. LegacyTranslate retrieves semantically similar PL/SQL→Java pairs with SFR-Mistral embeddings and uses $k = 3$ exemplars as the best trade-off between relevance and prompt length in experiments (Moti et al., 14 Mar 2026). RepoTransAgent separates retrieval from context acquisition: its RAG Agent queries a pair store and a name store, while its Context Agent iteratively invokes tools such as get_source_class_info, find_target_imports, and find_target_method_body to gather repository-specific facts in unified JSON (Guan et al., 25 Aug 2025).

Several systems replace raw history with structured memory. DelTA stores Proper Noun Records, Bilingual Summary, Long-Term Memory, and Short-Term Memory, and updates them sentence by sentence to improve consistency and avoid omissions (Wang et al., 2024). Loong generalizes this design into a 3E memory module—Essence, Exemplar, and Entity—and optimizes context selection from that memory with preference-based reinforcement learning (Wang et al., 28 May 2026). ViDove extends memory across modalities: its short-term memory stores translation history and multimodal cues for the current video, and its long-term memory stores domain knowledge and web retrieval results via Tavily, all accessed through LlamaIndex (Lu et al., 9 Jul 2025). CRAT narrows retrieval to flagged “unknown” or context-sensitive terms, then consolidates internal and external evidence into a Translation Knowledge Graph before allowing the Translator to consume it (Chen et al., 2024).

Grounding can also be stylistic or institutional rather than factual. SAMAS derives control signals from wavelet-packet statistics over word-length sequences, using Relative Wavelet Energy, Wavelet Entropy, skewness, and kurtosis to route texts through specialized literary agents (Wu et al., 23 Feb 2026). TransLaw grounds legal translation in Combined Department of Justice glossaries and in Proofreading Memory triplets of the form $\langle \text{src}, \text{ref}, \text{err} \rangle$ (Xuan et al., 1 Jul 2025). LegacyTranslate grounds code in an API knowledge base of 80 entries extracted from shared Java libraries (Moti et al., 14 Mar 2026). A recurring pattern is that translation agents increasingly externalize context into structured stores rather than relying on latent model memory alone.

4. Validation, feedback, and evaluation

Translation-agent systems are unusual in the centrality they give to validation. In code translation, compiler feedback, test execution, and structural checks are often direct inputs to subsequent agents. LegacyTranslate defines Structural Validity, Compilation Rate, and Test-Pass Rate, with $CR = C/N$ and $TPR = T/N$ (Moti et al., 14 Mar 2026). TRANSAGENT uses Computational Accuracy as the fraction of translated programs passing all tests, and narrows semantic repair by aligning source and target execution traces at block granularity (Yuan et al., 2024). RepoTransAgent treats compilation errors, runtime errors, functional errors, and non-terminating executions as distinct categories for reflection-driven repair (Guan et al., 25 Aug 2025). ReCodeAgent raises this to repository scale: across 118 projects, it reports 99.4% compilation success and improves validated developer-test pass rate by 60.8% over baselines (Ibrahimzada et al., 8 Apr 2026).

Natural-language translation work likewise shifts evaluation from single scalar metrics to multi-perspective analysis. TACTIC emphasizes XCOMET-XXL, MetricX-24-XXL, and COMETKIWI-23-XXL, and uses an internal ScoreAgent to aggregate faithfulness, expressiveness, and elegance into an acceptance score (Li et al., 10 Jun 2025). DITING argues that web-novel translation requires genre-specific evaluation across idiom translation, lexical ambiguity, terminology localization, tense consistency, zero-pronoun resolution, and cultural safety; its AgentEval debate framework achieves the highest correlation with human judgments among seven tested automatic metrics, with Spearman $\rho = 0.669$ for the multi-agent debate variant (Zhang et al., 10 Oct 2025). TransLaw uses a weighted legal-evaluation score

$I = \alpha A + \beta C + \gamma S$

with default weights $\alpha = 0.6$, $\beta = 0.3$, and $E_{\text{low}} > 0.6$0, where $E_{\text{low}} > 0.6$1 is legal semantic accuracy, $E_{\text{low}} > 0.6$2 is structural coherence and cohesion, and $E_{\text{low}} > 0.6$3 is stylistic appropriateness (Xuan et al., 1 Jul 2025).

Several papers explicitly question the adequacy of overlap-based metrics. CLASI introduces Valid Information Proportion, defined as $E_{\text{low}} > 0.6$4, to measure how much of a simultaneous interpretation is actually usable to the listener; on RealSI it reports VIP of 81.3% for Chinese-to-English and 78.0% for English-to-Chinese, while other systems achieve 35.4% and 41.6% (Cheng et al., 2024). TransAgents reports a d-BLEU of 25.0 yet wins 55.9% of bilingual LLM pairwise comparisons against GPT-4-1106-PREVIEW and 66.2% against a human reference, suggesting that literary preference and n-gram overlap can diverge sharply (Wu et al., 2024). DelTA formalizes proper-noun consistency with LTCR-1 and reports consistency gains of up to 4.58 percentage points (Wang et al., 2024). This suggests that translation agents are also forcing a redefinition of what counts as translation quality.

5. Domain-specific instantiations

The term “translation agent” now covers a wide spectrum of applied settings.

In enterprise code modernization, LegacyTranslate was developed in a financial-institution migration of approximately 2.5 million lines of PL/SQL to Java. The Initial Translation Agent alone achieved 45.6% compilable outputs and a 30.9% test-pass rate; the full pipeline raised these to 52.9% and 33.8% (Moti et al., 14 Mar 2026). RepoTransAgent, evaluated on hundreds of Java–C# pairs from six open-source repositories, achieved up to 55.34% compile rate and 45.84% pass rate, outperforming UniTrans, PLTranslation, and a no-agent baseline (Guan et al., 25 Aug 2025). TRANSAGENT targets general code translation across Python, Java, and C++, reaching Computational Accuracy as high as 94.5% depending on direction and outperforming UniTrans and TransCoder (Yuan et al., 2024).

In document and literary translation, DelTA improves consistency and COMET through multi-level memory while remaining sentence-wise and omission-resistant (Wang et al., 2024). Loong targets long-document translation with observe-and-act adaptive context selection and reports average gains of up to 13.0 points across sCOMET, dCOMET, and LLM-as-a-Judge (Wang et al., 28 May 2026). SAMAS focuses on style fidelity rather than only semantic correctness and reports FLORES-200 en→xx XCOMET of 96.93 when paired with GPT-5, with statistically significant gains over TACTIC (Wu et al., 23 Feb 2026). TransAgents simulates a literary publication workflow with CEO, Senior Editor, Junior Editor, Translator, Localization Specialist, and Proofreader roles for ultra-long literary texts (Wu et al., 2024).

In multimodal and streaming translation, ViDove integrates auditory and visual agents with multimodal memory for video subtitling, achieving BLEU 23.51 and SubER 73.38 on DoveBench, improving BLEU by 28.8% and reducing SubER by 15.5% over Whisper + DelTA (Lu et al., 9 Jul 2025). CLASI treats simultaneous speech translation as an agentic read–retrieve–plan–write loop and reports 70% VIP on an extremely hard dataset where other systems remain under 13% (Cheng et al., 2024). Agent-SiMT separates policy selection from translation generation by pairing an HMT-based policy agent with an LLM translation agent, yielding state-of-the-art latency–quality trade-offs in simultaneous machine translation (Guo et al., 2024).

In tightly constrained domains, specialized translation agents handle structure or regulation directly. LaTeXTrans uses Parser, Translator, Validator, Summarizer, Terminology Extractor, and Generator agents to preserve compilability and structure in LaTeX documents; on En–Zh with a GPT-4o backbone it reports COMETkiwi 73.59, LLM-score 8.92, and FC-score 71.52 (Zhu et al., 26 Aug 2025). TransLaw adapts the pattern to Hong Kong legal judgments and reports that its TransLaw-ChatGPT configuration improves legal semantic accuracy, structural coherence, and stylistic fidelity over GPT-4o, with ACS gains across all tested weightings (Xuan et al., 1 Jul 2025).

6. Limitations, controversies, and research directions

Despite strong results, current work presents a consistent set of limitations. First, multi-agent decomposition is not free. TACTIC notes that multi-stage workflows increase inference time and cost relative to single-pass translation (Li et al., 10 Jun 2025). Loong requires trajectory sampling for preference construction and uses 4×A800 GPUs for roughly 20 hours of training per model (Wang et al., 28 May 2026). LegacyTranslate reports that large prompts and iterative refinement incur computational cost, with experiments using an A100 GPU for about five hours total (Moti et al., 14 Mar 2026). ReCodeAgent averages about $15.3 and about 57 minutes per project (Ibrahimzada et al., 8 Apr 2026).

Second, retrieval and memory can help or harm. CRAT explicitly argues that straightforward RAG can overload the model and induce hallucinations, motivating targeted retrieval and causal validation (Chen et al., 2024). DelTA and Loong both treat redundant context as actively harmful, not merely unnecessary (Wang et al., 2024, Wang et al., 28 May 2026). RepoTransAgent shows that smaller models may see smaller gains or even declines because complex prompts and tool invocation become difficult to follow (Guan et al., 25 Aug 2025). LegacyTranslate states that retrieval quality and API-knowledge-base completeness are pivotal and that hallucinations remain a risk for framework glue code without clear examples (Moti et al., 14 Mar 2026).

Third, validation itself remains imperfect. ACToR demonstrates that adversarial testing materially improves C→Rust correctness, but it still depends on discriminating test suites and assumes deterministic, single-threaded programs (Li et al., 4 Oct 2025). CLASI argues that BLEU, BLEURT, and COMET correlate poorly with human VIP at higher quality (Cheng et al., 2024). MAATS improves true-positive error detection substantially but also increases false positives, though the reported excess is mostly minor (Wang et al., 20 May 2025). TransLaw observes that multi-turn revision can induce hallucinations and that contextualizing complex legal terminology and achieving stylistic naturalness still trail human experts (Xuan et al., 1 Jul 2025).

Current research directions therefore converge on a few themes. One is richer grounding: curated termbases, API knowledge bases, multimodal evidence, and legal or domain-specific retrieval (Moti et al., 14 Mar 2026, Lu et al., 9 Jul 2025, Xuan et al., 1 Jul 2025). Another is better control over long context, through compact summaries, adaptive retrieval, and learned routing policies (Wang et al., 2024, Wang et al., 28 May 2026). A third is stronger validation, including execution alignment, adversarial testing, coverage-guided validation, and repository-scale trajectory analysis (Yuan et al., 2024, Li et al., 4 Oct 2025, Ibrahimzada et al., 8 Apr 2026). A plausible implication is that the future translation agent will be less a standalone translator than a validated, memory-bearing controller over translation-specific tools, evaluators, and external knowledge.