SAMAS: A Spectrum-Guided Multi-Agent System for Achieving Style Fidelity in Literary Translation

Jingzhuo Wu; Jiajun Zhang; Keyan Jin; Dehua Ma; Junbo Wang

cs.CLFeb 23, 2026

SAMAS: A Spectrum-Guided Multi-Agent System for Achieving Style Fidelity in Literary Translation

Authors: Jingzhuo Wu, Jiajun Zhang, Keyan Jin, Dehua Ma, Junbo Wang

Organizations: Beijing Normal University · University of Science and Technology of China · University of Coimbra · Beijing University of Posts and Telecommunications ·

Abstract

Modern large language models (LLMs) excel at generating fluent and faithful translations. However, they struggle to preserve an author's unique literary style, often producing semantically correct but generic outputs. This limitation stems from the inability of current single-model and static multi-agent systems to perceive and adapt to stylistic variations. To address this, we introduce the Style-Adaptive Multi-Agent System (SAMAS), a novel framework that treats style preservation as a signal processing task. Specifically, our method quantifies literary style into a Stylistic Feature Spectrum (SFS) using the wavelet packet transform. This SFS serves as a control signal to dynamically assemble a tailored workflow of specialized translation agents based on the source text's structural patterns. Extensive experiments on translation benchmarks show that SAMAS achieves competitive semantic accuracy against strong baselines, primarily by leveraging its statistically significant advantage in style fidelity.

Figures & tables

Figure 1: Overview of the SAMAS. The system first quantifies a text’s literary style into a Stylistic Feature Spectrum (SFS) signal. This signal then guides the dynamic allocation of a workflow, which is executed by a pool of specialized agents to produce a stylistically faithful translation that outperforms a single-LLM baseline.

Single-LLM
Models	Routing Method	FLORES-200				WMT24
		en $\rightarrow$ xx		xx $\rightarrow$ en		en $\rightarrow$ xx		xx $\rightarrow$ en
		XC	KW	XC	KW	XC	KW	XC	KW
Qwen3-235B-A22B [ 14 ]	-	84.17	78.25	94.05	87.17	73.97	68.57	82.59	77.83
Qwen3-30B-A3B [ 14 ]	-	77.53	70.18	91.55	85.12	66.97	59.97	81.21	77.11
Qwen-mt-plu [ 14 ]	-	86.97	81.27	94.43	87.43	76.21	70.87	84.37	79.40

Table 1: Comparison of translation performance on the FLORES-200 and WMT24 benchmarks, evaluated with XCOMET (XC) and COMETKIWI-23 (KW) metrics. The “Models” column specifies the base LLM used for translation, while the “Routing Method” column details the workflow applied within our SAMAS framework (e.g., dynamic routing or fixed style-based routes). “Qwen3-235B-A22B” denotes the Instruct version. The best and second-best scores are highlighted. Scores marked with * indicate a statistically significant improvement over the strongest multi-agent baseline (TACTIC) with p

<

0.05.

Figure 2: Human preference for style fidelity in head-to-head comparisons. In all evaluations, translations generated by SAMAS were consistently favored by human evaluators over those from five different strong baseline models.

Machine Translation

Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.

See all 267 papers

AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing
Mar 24, 2026Sarubi Thillainathan, Ji-Ung Lee, Michael Sullivan +1Text Style Transfer Text Generation
Department of Language Science and Technology, Saarland Informatics Campus Saarland University, Saarbrücken, Germany
Worlds Within Words: Translating Culture in Ancient Chinese Texts with Multi-Agent Coordination
May 31, 2026Xiaoqi He, Kaixin Lan, Mu You +3Multi-Agent LLMs Machine Translation
NLP2CT Lab, Department of Computer and Information Science, University of Macau · Institute of International Language Services Studies, Macau Millennium College
Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach
Jun 4, 2026Zhihao Lin, Ziqi Zhu, Hao Huang +2Synthetic Data Generation for Language Models Language Model Post-Training
1Amazon Web Services (AWS) · 2Peking University

Mar 24, 2026J/K move · Enter open · S save

AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing

Sarubi Thillainathan, Ji-Ung Lee, Michael Sullivan, Alexander Koller

Department of Language Science and Technology, Saarland Informatics Campus Saarland University, Saarbrücken, Germany

Text Style Transfer Text Generation

Open paper AskPDF arXiv

The task of authorship style transfer involves rewriting text in the style of a target author while preserving the meaning of the original text. Existing style transfer methods train a single model on large corpora to model all target styles at once: this high-cost approach offers limited flexibility for target-specific adaptation, and often sacrifices meaning preservation for style transfer. In this paper, we propose AuthorMix: a lightweight, modular, and interpretable style transfer framework. We first train individual, style-specific LoRA adapters on a small set of high-resource authors: this allows for the rapid training of specialized adaptation models for each new target using layer-wise adapter mixing via reinforcement learning, necessitating only a handful of target-style training examples. AuthorMix ranks first on the combined style-meaning score among all baselines, including GPT-5.1, and substantially improves meaning preservation over the trained baselines; under human evaluation it is the only method best-or-tied on every dimension.

Abstract

Machine Translation

Explore similar work

AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing

Worlds Within Words: Translating Culture in Ancient Chinese Texts with Multi-Agent Coordination

Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach