Paper deep dive
Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation
Hasan Alkhder, Mohammad Abboush, Igor Tchappi, Ahmet Zengin, Amro Najjar
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/4/2026, 4:29:19 AM
Summary
This paper proposes reframing neural machine translation (NMT) as a structured decision space explored by autonomous translation agents to address opacity in low-resource dialect generation. Using Turkish-Syrian Arabic translation as a case study, the authors model three distinct translation pathways (agents) operating over a shared multilingual backbone: zero-shot direct translation, dialect-stabilized translation via lightweight fine-tuning, and pivot translation through English. The study evaluates these agents on 5,000 dialogue sentences, quantifying divergence using dialect marker frequency, lexical proximity to Modern Standard Arabic, and structural variance. Results indicate that lightweight stabilization nearly doubles dialect marker usage and reduces structural instability, while pivot mediation introduces normalization pressure. The framework treats inter-agent divergence as an interpretable behavioral signal rather than error, providing a principled approach for analyzing low-resource dialect generation.
Entities (10)
Relation Signals (8)
Hasan Alkhder → authored → Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation
confidence 99% · Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation Hasan Alkhder
Turkish-Syrian Arabic Translation → usesmethod → Dialect-Stabilized Translation
confidence 95% · We conduct an empirical study on Turkish--Syrian Arabic translation using three agents: (2) dialect-stabilized translation via lightweight fine-tuning
Turkish-Syrian Arabic Translation → usesmethod → Pivot Translation
confidence 95% · We conduct an empirical study on Turkish--Syrian Arabic translation using three agents: (3) pivot translation through English
Turkish-Syrian Arabic Translation → usesmethod → Zero-shot Direct Translation
confidence 95% · We conduct an empirical study on Turkish--Syrian Arabic translation using three agents: (1) zero-shot direct translation
Dialect-Stabilized Translation → improvesmetric → Dialect Marker Frequency
confidence 94% · Lightweight stabilization nearly doubles dialect marker usage, increasing it from 0.2266 to 0.4988
Dialect-Stabilized Translation → reducesmetric → Structural Variance
confidence 93% · Lightweight stabilization nearly doubles dialect marker usage... while significantly reducing structural instability
Dialect-Stabilized Translation → trainedon → MADAR-Turk
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Neural machine translation (NMT) systems typically produce a single output per input, obscuring the alternative decision trajectories implicitly available within multilingual decoding. This opacity becomes particularly problematic in low-resource dialect settings, where multiple linguistically valid realizations may differ in lexical authenticity, register, and structural stability. We propose reframing translation as a structured decision space explored by autonomous translation agents. Instead of analyzing a single output, we model distinct translation pathways as agents operating over a shared multilingual backbone. Inter-agent divergence is treated not as error but as an interpretable behavioral signal. We conduct an empirical study on Turkish--Syrian Arabic translation using three agents: (1) zero-shot direct translation, (2) dialect-stabilized translation via lightweight fine-tuning, and (3) pivot translation through English. Evaluation is performed on 5,000 dialogue sentences, while stabilization is trained on 5,000 additional Turkish--Syrian sentence pairs drawn from television dialogue and MADAR-Turk resources. Rather than optimizing for conventional performance metrics, we quantify structured behavioral displacement using dialect marker frequency, lexical proximity to standardized Arabic, and structural variance. Lightweight stabilization nearly doubles dialect marker usage, increasing it from 0.2266 to 0.4988, while significantly reducing structural instability. Pivot mediation introduces normalization pressure and measurable compression effects, whereas zero-shot translation exhibits the highest decision variance. We argue that translation divergence across agents reveals latent decision flexibility within multilingual models and we provide a principled interpretability framework for low-resource dialect generation.
Tags
Links
- Source: https://arxiv.org/abs/2609.04048v1
- Canonical: https://arxiv.org/abs/2609.04048v1
Trouble viewing inline? Open PDF directly →
Full Text
45,479 characters extracted from source content.
Expand or collapse full text
Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation Hasan Alkhder 1⋆ Mohammad Abboush 2 Igor Tchappi 3 Ahmet Zengin 1 Amro Najjar 4 AbstractNeural machine translation (NMT) systems typically produce a single output per input, obscuring the alternative decision trajectories implicitly available within multilingual decoding. This opacity becomes particularly problematic in low-resource dialect settings, where multiple linguistically valid realizations may differ in lexical authenticity, register, and structural stability. We propose reframing translation as a structured decision space explored by autonomous translation agents. Instead of analyzing a single output, we model distinct translation pathways as agents operating over a shared multilingual backbone. Inter-agent divergence is treated not as error but as an interpretable behavioral signal. We conduct an empirical study on Turkish–Syrian Arabic translation using three agents: (1) zero-shot direct translation, (2) dialect-stabilized translation via lightweight fine-tuning, and (3) pivot translation through English. Evaluation is performed on 5,000 dialogue sentences, while stabilization is trained on 5,000 additional Turkish–Syrian sentence pairs drawn from television dialogue and MADAR-Turk resources. Rather than optimizing for conventional performance metrics, we quan- tify structured behavioral displacement using dialect marker frequency, lexical proximity to standardized Arabic, and structural variance. Lightweight stabilization nearly doubles dialect marker usage, increasing it from 0.2266 to 0.4988, while significantly reducing structural instability. Pivot mediation introduces normalization pressure and measurable compres- sion effects, whereas zero-shot translation exhibits the highest decision variance. We argue that translation divergence across agents reveals latent deci- sion flexibility within multilingual models. By modeling translation as a multi-agent system over a shared decision space, we provide a principled interpretability framework for low-resource dialect generation. Keywords:Explainable AI, Multi-Agent Systems, Dialect Translation, Decision Space Modeling, Low-Resource NLP ⋆ Corresponding author: hasan.alkhder2@ogr.sakarya.edu.tr 1 arXiv:2609.04048v1 [cs.CL] 3 Sep 2026 2 1 Introduction Neural machine translation (NMT) systems are commonly treated as determin- istic functions that map a source sentence to a single target realization. In practice, however, multilingual generative models operate over high-dimensional probability distributions that encode multiple plausible translation trajectories. Despite this inherent flexibility, only one output surface form is typically exposed to the user. The alternative decisions embedded within the decoding space re- main hidden. This creates a limitation for dialect translation analysis, because multiple valid realizations may exist while standard translation pipelines expose only a single output. This opacity becomes especially consequential in low-resource dialect set- tings. Unlike standardized language translation, dialect translation involves competing realizations that differ not only lexically but socially and stylisti- cally. Dialectal outputs may vary in register, authenticity, emotional intensity, and structural compression. In such contexts, divergence across outputs is not necessarily error; it may reflect meaningful trade-offs within the model’s inter- nal decision landscape. Existing MT evaluation approaches capture translation adequacy reasonably well, but they are less suited for analyzing dialect authen- ticity, register choice, and structural variation across alternative translation re- alizations. Syrian Arabic provides a particularly challenging case. It lacks standardized orthography, exhibits lexical divergence from Modern Standard Arabic (MSA), and is supported by limited parallel resources [6]. Multilingual models trained primarily on standardized data frequently display instability when generating dialectal text. This instability manifests as partial normalization toward MSA, repetition collapse, structural drift, or inconsistent colloquial weighting. Metrics such as BLEU [18] and COMET [19], while effective for measuring translation adequacy, do not capture dialectal register, normalization pressure, or structural behavioral differences across translation pathways. In this work, we propose reframing translation as adecision spaceexplored by translation agents. We adopt an operational definition of agents as constrained decoding pathways within a shared multilingual model, rather than fully au- tonomous entities in the classical multi-agent systems (MAS) sense. Each agent corresponds to a specific decision constraint applied to the same underlying model. Rather than treating translation as a single-output mapping, we model distinct translation pathways as independent agents operating over a shared multilingual backbone. Each agent embodies a specific decision constraint— zero-shot generation, dialect-reinforced adaptation, or pivot mediation—while preserving identical architecture and decoding parameters. We use the term agent because each pathway produces systematically different translation be- havior under a defined constraint regime. The observable divergence between agents is interpreted not as noise, but as structured behavioral signal. We conduct an empirical case study on Turkish–Syrian Arabic translation. Three agents are constructed: (1) direct zero-shot Turkish→Syrian translation, (2) dialect-stabilized translation via lightweight fine-tuning on 5,000 Turkish– 3 Syrian sentence pairs, and (3) pivot translation through English (Turkish→ English→Syrian). Evaluation is performed on 5,000 aligned dialogue segments drawn from independent dialogue sources. This controlled setup allows us to isolate decision-space reweighting effects from architectural changes. Instead of optimizing for conventional performance metrics, we quantify inter-agent divergence using dialect marker frequency, lexical proximity to stan- dardized Arabic, and structural length variance. Together, these metrics capture complementary aspects of dialect activation, normalization pressure, and struc- tural stability across translation pathways. Our findings show that lightweight dialect stabilization nearly doubles dialect marker usage while reducing struc- tural instability, whereas pivot mediation introduces compression and normal- ization effects. Zero-shot translation exhibits the highest variance, reflecting under-specified dialect representation. By modeling translation as a multi-agent system operating within a shared decision space, we introduce an interpretability-oriented framework for analyz- ing low-resource dialect generation. Rather than evaluating translation solely through final-output quality, our approach examines how controlled variation in translation pathways reveals structured behavioral differences within the same multilingual model. This perspective reframes divergence not as translation fail- ure, but as an observable signal of how multilingual systems balance authenticity, normalization, and semantic trade-offs under different decision constraints. The central research question of this work is whether controlled translation path- way variation can reveal systematic and interpretable differences in low-resource dialect translation behavior. Importantly, this study does not aim to improve machine translation ac- curacy. Instead, translation is used as a controlled generative environment in which agent-level behavioral divergence can be observed and analyzed. The contributions of this work are threefold. First, we introduce a decision- space perspective for analyzing multilingual translation systems by modeling translation pathways as constrained agents. Second, we propose behavioral met- rics to quantify inter-agent divergence in dialect activation, lexical normalization, and structural realization. Third, we provide an empirical analysis of Turkish– Syrian Arabic translation demonstrating how stabilization and pivot mediation reshape decoding behavior. We emphasize that the notion of ”agents” in this work is an abstract and analytical construct rather than a realization of classical multi-agent systems. The objective is not to model autonomous or interacting agents, but to provide a structured perspective for analyzing controlled decoding pathways within a shared multilingual model. In this sense, agents represent constrained opera- tional views of the same system, enabling systematic exploration of the transla- tion decision space. 4 2 Related Work Transfer Learning for Low-Resource MTTransfer learning has become a widely adopted strategy for improving translation performance in data-scarce scenar- ios. Several studies have shown that knowledge acquired from high-resource language pairs can be transferred to low-resource tasks. For instance, Neubig et al. proposed similar-language regularization through joint training on related languages to reduce overfitting [17]. Gu et al. framed low-resource translation as a meta-learning problem that enables rapid adaptation from limited data [9]. Kim et al. investigated cross-lingual transfer using language-agnostic encoders and cross-lingual embeddings to alleviate vocabulary mismatch [10], while Luo et al. introduced a hierarchical transfer framework that progressively trains on unrelated high-resource pairs, intermediate languages, and finally low-resource pairs [14]. Although these techniques significantly improve translation quality, they gen- erally produce deterministic single-output translations without revealing alter- native translation hypotheses or underlying decision mechanisms. Arabic Dialect TranslationArabic dialect translation introduces additional chal- lenges due to morphological complexity, code-switching, and the limited avail- ability of dialectal corpora. In this context, Sibaee et al. proposed SHAMI- MT, a bidirectional translation system for Syrian Arabic and Modern Standard Arabic based on AraT5v2 and trained on the Nabra dataset [22]. Baniata et al. developed a multitask learning architecture that shares a decoder across dialect translation tasks to exploit cross-task information [3]. Salloum et al. demonstrated that unsupervised morphological segmentation can improve di- alectal translation performance [20], while Shapiro et al. examined the influence of dialect identification on Arabic neural machine translation [21]. Furthermore, the MADAR shared tasks have advanced dialect identification research, achieving macro-averaged F1-scores above 67% across 25 dialects [1, 23]. Alkheder et al. introduced the MADAR-Turk dataset, which contains man- ually translated Damascus–Turkish sentence pairs intended to facilitate bench- marking in dialectal translation tasks [2]. Despite these developments, most approaches treat translation as a single- stage mapping process and provide limited interpretability regarding the influ- ence of lexical and morphological variations on translation decisions. Explainability in Neural Machine TranslationExplainable AI techniques have also been explored to analyse the behaviour of neural translation models. Fer- rando et al. proposed ALTI+, which traces token-level attribution across source and target representations [8]. Voita et al. adapted Layerwise Relevance Propa- gation for Transformer models to measure source and target contributions [24]. Moradi et al. showed that different attention distributions can produce iden- tical predictions, highlighting the limitations of attention-based explanations [15]. Additional studies have evaluated explanation fidelity and attention-head importance in translation models [25]. 5 However, these approaches mainly provide post-hoc interpretations of single outputs rather than exposing the broader decision processes underlying transla- tion generation. Multi-Agent Approaches and Pivot TranslationRecent work has also explored multi-agent frameworks for machine translation. Wu et al. introduced TransAgents, a collaborative architecture inspired by human translation workflows [26]. Other studies proposed multi-agent learning strategies based on ensemble distillation, mutual learning, or collaborative debate mechanisms [5,13,12]. While these methods improve translation quality through collaborative re- finement, they typically treat agents as separate models rather than as entities exploring structured translation alternatives. Pivot translation represents another approach for low-resource language pairs by exploiting high-resource intermediate languages. For example, Kim et al. proposed pivot-based transfer learning through step-wise pre-training [11]. Ad- ditional work investigated multi-pivot ensembling and simultaneous multi-pivot translation strategies to improve robustness [16,7]. Despite their effectiveness, pivot-based approaches may propagate errors across translation stages and provide limited insight into the decision processes involved in pivot-mediated translation. Summary and Research GapOverall, the existing literature reveals an impor- tant research gap in neural machine translation for low-resource dialects. Most approaches focus on improving translation quality while treating translation as a deterministic mapping that generates a single output. Consequently, current methods offer limited insight into the internal decision processes of translation models or the alternative translation trajectories that may exist. This limitation is particularly critical for dialectal translation, where linguistic variability and limited data availability require more interpretable systems capable of explaining how lexical and structural choices are made during translation. 3 Data Resources and Experimental Setup 3.1 Evaluation Dataset The evaluation corpus consists of 5K sentence-level dialogue segments aligned across multiple linguistic representations. The dataset is constructed as a fixed evaluation set rather than a benchmark-scale testbed. Since the goal of this study is controlled behavioral comparison between translation pathways, a stable set of 5K sentences provides sufficient coverage while enabling systematic analysis under identical conditions. All translation agents operate on the same evaluation corpus so that observed behavioral differences can be attributed to translation pathway effects rather than dataset variation. Each entry contains four aligned linguistic representations. The Turkish col- umn (TR) contains the original source sentence. The Syrian Arabic column (SY) 6 provides the reference dialectal translation derived from professionally dubbed dialogue. The English column (EN) contains an aligned English rendering used for pivot translation modeling. The Modern Standard Arabic column (MSA) provides a standardized Arabic form used for lexical proximity comparison. Although the agents generate Turkish→Syrian outputs exclusively, the pres- ence of English and MSA columns enables structured analysis of pivot mediation and normalization effects. The 5K evaluation sentences are drawn from two complementary sources. The first consists of 3K dialogue segments extracted from a professionally dubbed Turkish television series provided by NIS Studios, capturing naturally occurring conversational speech with emotional and informal discourse patterns. The second consists of 2K conversational sentence pairs from the MADAR- Turk corpus [2], providing structurally controlled conversational data. Combining these sources balances natural dialogue richness with structurally controlled conversational data. Television dialogue contributes authentic collo- quial phrasing and discourse markers, while the additional conversational data provides controlled coverage of common conversational patterns. Importantly, the evaluation dataset is constructed independently and does not overlap with the stabilization (fine-tuning) data, ensuring that all reported results reflect generalization rather than memorization. Example Entry. TR:Ben ümitli değilim açıkçası. EN:I’m not hopeful, to be honest. MSA:أؔ ۹ݐ ܇ـوٴ؇ே ศ༥ؑ ௌًاڶ. SY:؊ש أؔ ܇ᇖᆇ ۑـ૯ે ܇ـوٴ؇ჱჩ ؑঔঐݮܰݮع. This multi-column alignment enables relational comparison between dialectal output, standardized Arabic, and pivot-mediated translations without constrain- ing agents to a single reference target. 3.2 Dialect Stabilization Dataset To model dialect reinforcement as decision-space reweighting, we construct a separate fine-tuning dataset consisting of 5K Turkish–Syrian sentence pairs. The stabilization corpus combines two sources as shown in Table 1. The television dialogue portion was obtained from professionally dubbed Turkish television scripts provided by NIS Studios, a media production studio specialized in Turkish–Syrian Arabic dubbing. The MADAR-Turk corpus is used entirely for stabilization (fine-tuning), and no internal split is applied. This composition enables effective dialect reinforcement while maintaining diversity in conversational structure and avoiding overfitting to a single narrative domain. 7 Table 1: Dialect Stabilization Dataset Composition SourceNumber of Sentences Television Dialogue (NIS)3K MADAR-Turk Conversational Pairs2K Total5K 3.3 Fine-Tuning Configuration Fine-tuning is intentionally lightweight. The objective is not to maximize end- task translation performance but to introduce controlled dialectal reinforcement within the existing multilingual representation space. The model is fine-tuned on 5K Turkish–Syrian sentence pairs for a single epoch using a learning rate of1×10 −5 . No vocabulary extension or architectural modifications are applied, and all agents share identical decoding parameters. This constrained training regime limits parameter updates substantially, en- suring that observed behavioral shifts primarily reflect lexical reweighting rather than large-scale structural retraining. 3.4 Experimental Control and Separation Strict separation between evaluation and stabilization datasets is enforced. Al- though both datasets are constructed from similar sources (NIS dialogue and MADAR-Turk), they consist of distinct sentence instances with no overlap at the sentence level. By controlling architecture, decoding parameters, and dataset separation, translation pathway remains the only varying factor across agents. This con- trolled setup enables systematic comparison of translation behavior within the proposed decision-space framework. 4 Multi-Agent Translation Framework Neural machine translation systems are often described as mappings of the form f(x) =y, where a source sentencexproduces a target outputy. This represen- tation is a simplifying abstraction. In practice, multilingual generative models define a conditional probability distribution over possible outputs. For a given inputx, decoding explores a high-dimensional distribution that contains mul- tiple plausible realizations. The single output returned to the user therefore represents only one trajectory through a broader latent decision landscape. We interpret this flexibility as a structured translation decision space. In this work, the decision space refers to the set of outputs generated when the same multilingual model is executed under different controlled pathway constraints. These outputs represent observable samples from the model’s conditional output distribution. 8 Figure1: Overview of the proposed decision-space framework. A single multilingual model is executed under three constrained translation pathways (agents): zero-shot, stabilized, and pivot-based. The resulting outputs represent different regions of the shared translation decision space. As illustrated in Figure 1, translation pathways define distinct regions within the shared decision space. Formally, for a Turkish input sentencex, the multilingual backbone model defines a conditional distributionP(y|x)over possible Syrian Arabic realiza- tions. Rather than treating this distribution as an opaque sampling mechanism, we introduce controlled translation pathways that modify how decoding explores the distribution. Each pathway is modeled as a translation agent representing a constrained decoding configuration of the same underlying model. We emphasize that agents in this framework are not fully autonomous entities in the classical MAS sense, but operational configurations defined by distinct decision constraints applied to a shared multilingual system. LetMdenote the fixed multilingual model. An agentA i maps an input sentencexto an outputy i under a specific operational constraintC i : A i (x) =M(x;C i ). The constraintC i does not modify the architecture ofMbut specifies the translation pathway under which decoding is performed. To analyze how translation behavior changes under different operational con- ditions, we examine three representative pathway configurations. 9 The first agent performs direct zero-shot Turkish→Syrian translation using the pretrained multilingual representation without dialect-specific adaptation. The second agent introduces dialect stabilization through lightweight fine- tuning on 5K Turkish–Syrian sentence pairs. The fine-tuning is intentionally minimal so that the multilingual representation is not substantially retrained, but lexical activation patterns encouraging dialectal realization are slightly re- inforced. The third agent implements pivot mediation, decomposing translation into two sequential mappings: Turkish→English followed by English→Syrian. The intermediate English representation acts as a semantic bottleneck that can in- troduce additional normalization pressure on the final output. All agents share identical decoding parameters and operate on the same evaluation dataset. The only variation lies in the pathway constraintC i . For each inputx, we observe a subset of the translation decision space gen- erated by the selected pathway configurations: D(x) =A 1 (x),A 2 (x),A 3 (x). This set represents observable realizations under the three controlled pathway constraints rather than the full space of outputs defined byP(y|x). The elements ofD(x)are not interpreted hierarchically as correct versus in- correct translations. Instead, they represent neighboring realizations within the same semantic manifold. Divergence between agents therefore reflects behavioral displacement within the model’s latent representation space. This perspective shifts interpretability from internal probing toward rela- tional analysis of outputs. Instead of analyzing hidden representations directly, we observe how translations change when decoding conditions are modified in a controlled manner [4]. In this framework, divergence across agents becomes an interpretable signal revealing how multilingual models balance dialectal authen- ticity, normalization pressure, and structural stability. Importantly, the agents are not independent models but coordinated decod- ing configurations of the same multilingual system. The multi-agent perspective therefore introduces structured pathway variation rather than additional archi- tectures. Within low-resource dialect settings, this variation is particularly informa- tive. Instability in zero-shot generation, normalization under pivot mediation, and reinforcement under dialect stabilization correspond to different observable regions of the translation decision space, allowing systematic analysis of dialect generation behavior without modifying the underlying model architecture. 5 Behavioral Metrics for Structured Divergence Standard reference-based metrics remain useful for measuring translation ade- quacy. However, they are insufficient for capturing behavioral variation across translation pathways. In this work, conventional correctness-based evaluation is therefore complemented with behavior-oriented analysis that measures how 10 translation outputs shift under different decoding conditions. Instead of eval- uating a single output against a reference, we quantify structured divergence within the translation decision space introduced in Section 4. These metrics are designed to provide interpretable signals at the behav- ioral level. Dialect marker frequency reflects lexical-level interpretability, lexical overlap captures normalization tendencies, and structural length variation re- flects stability in generation behavior. The selected metrics capture three complementary aspects of translation be- havior: dialectal lexical activation, normalization toward standardized Arabic, and structural realization. Letxdenote a Turkish source sentence andy i =A i (x)the output of agent A i . 5.1 Dialect Marker Frequency LetMdenote a predefined set of Syrian dialect markers curated from the dia- logue corpus and validated by native speakers (e.g.,݈ݮ، ۑـ૯ે، ڷسڿ، ܇ݮ، ۹ݐ). Dialect activation for a sentenceyis defined as: D(y) = 1 |y| X t∈y 1(t∈M) Corpus-level activation for agentA i over datasetXis: D A i = 1 |X| X x∈X D(A i (x)) Higher values indicate stronger dialect reinforcement in the generated trans- lations. 5.2 Lexical Proximity to Standardized Arabic Lexical proximity to Modern Standard Arabic (MSA) is used to estimate normal- ization pressure in the translation process. Since Syrian dialect diverges lexically from MSA, stronger overlap with standardized Arabic may indicate a tendency toward normalization rather than dialectal realization. Given standardized referenceS(x), lexical overlap is defined as: O(y,S) = |tokens(y)∩tokens(S(x))| |tokens(S(x))| Agent-level proximity is computed as: O A i = 1 |X| X x∈X O(A i (x),S(x)) Lower overlap therefore indicates stronger dialectal divergence from stan- dardized Arabic. 11 5.3 Structural Length Ratio Structural realization is measured through relative sentence length compared with the standardized reference. L(y) = ℓ(y) ℓ(S(x)) Corpus mean for agentA i : L A i = 1 |X| X x∈X L(A i (x)) Structural variance is computed as: σ L,A i = s 1 |X| X x∈X (L(A i (x))−L A i ) 2 Higher variance indicates instability in structural realization across generated translations. 5.4 Inter-Agent Divergence To quantify behavioral displacement between agents, we measure divergence across the three behavioral dimensions introduced above: dialect activation, lexical normalization, and structural realization. Relational displacement between agentsA i andA j is defined as: ∆ ij = D A i −D A j + O A i −O A j + L A i −L A j This score measures behavioral divergence between translation agents under different pathway constraints. The three components are combined with equal weight because they cap- ture complementary aspects of translation behavior and operate on different behavioral dimensions rather than comparable numeric scales. Equal weighting therefore avoids privileging one behavioral dimension over the others. 6 Quantitative Results Table 2 reports aggregate behavioral metrics across the three translation agents. Because all agents share the same multilingual backbone, observed differences reflect pathway-induced decision reweighting rather than architectural variation. To assess the robustness of observed differences, we performed paired com- parisons across agents using bootstrap resampling over the evaluation set. The observed differences in dialect marker frequency and length ratios remained sta- ble across resampled subsets, indicating that the reported trends are not driven by dataset-specific effects. 12 Table 2: Aggregate Behavioral Metrics Across Translation Agents MetricZero-shot Stabilized Pivot Avg. Dialect Markers 0.2266 0.4988 0.2578 Avg Overlap with MSA 0.2133 0.1117 0.1114 Std Overlap0.2530 0.1717 0.1788 Avg Length Ratio1.5498 1.0332 0.9354 Std Length Ratio2.8761 1.1834 1.0669 Dialect stabilization produces the strongest behavioral shift. Average dialect marker frequency increases from 0.2266 in zero-shot translation to 0.4988 af- ter fine-tuning, indicating substantial lexical reweighting toward Syrian dialect forms. At the same time, structural variance decreases sharply (σ L = 2.8761→ 1.1834), demonstrating improved stability in sentence realization. Lexical over- lap with standardized Arabic decreases (0.2133 → 0.1117), confirming stronger dialect divergence from MSA. The pivot pathway exhibits a different profile. Dialect activation remains moderate (0.2578), while the average length ratio drops below unity (0.9354), indicating structural compression relative to standardized references. Such com- pression often reflects normalization effects introduced by pivot mediation, where colloquial expansions or emphatic constructions in Syrian dialect are reduced when passing through an intermediate English representation. Variance is re- duced compared to zero-shot but remains slightly higher than stabilized, consis- tent with semantic mediation through English acting as a regularizing constraint. As illustrated in Figure 2, the three agents occupy distinct regions within the behavioral space. Zero-shot translation exhibits high variance and limited dialect activation. Stabilization shifts the system toward stronger dialect re- inforcement and structural consistency. Pivot translation compresses output length and moderates lexical activation, revealing measurable behavioral differ- ences consistent with decision-space reconfiguration across translation pathways. Importantly, qualitative inspection of outputs indicates that increased di- alectal activation does not appear to systematically degrade semantic adequacy. While minor lexical deviations occur, the core meaning is generally preserved across agents. 7 Discussion: Decision-Space Reconfiguration Across Agents The empirical results reveal systematic behavioral displacement across transla- tion agents. These shifts are not stochastic variations but structured reconfigura- tions within a shared multilingual decision space. Because all agents operate over the same backbone architecture, observed differences reflect pathway-induced reweighting rather than architectural divergence. 13 Table 3: Representative Sentence-Level Divergence Across Translation Agents Example 1 TR:Nasıl olacak? Nasıl yaşayacağız böyle? A1 Zero-shot:ۑسٞ رح ؊ۣݮن؟ ۑسٞ ݅ؽשཛྷܽ ڷიჀا؟ A2 Stabilized:ۑسٞ ౼ۣݮن؟ ۑسٞ ౼שཛྷݓ൘ഽ ڷسڿ؟ A3 Pivot:ۑسٞ ݅ؽשཛྷܽ ڷიჀا؟ Observed Pattern:Register Shift + Addressee Reframing Example 2 TR:Hadi açın kitaplarınızı. Müzik Müzik Müzik... A1 Zero-shot:ا٩ـڢݮا ۑـغઅ. ܇سݮز؊અ ܇سݮز؊અ... A2 Stabilized:ۿسؽ ا٩ـڢݮا ۑـغઅ A3 Pivot:ۿسؽ ؔوٴـڀ ۑـ اঔঐݮ݅سߚ߫... Observed Pattern:Token Noise + Contextual Normalization Example 3 TR:Neredeydi şu bardaklar? A1 Zero-shot:و੪ ڷسڿ ا۹ـۣݮب؟ A2 Stabilized:وಕ౿ၨထ ੪ ا۹ـۣݮؑت؟ A3 Pivot:وಕ౿ၨထ ੪ ا۹ؽݟرات؟ Observed Pattern:Lexical Drift / Semantic Misalignment Example 4 TR: O teyze benim yüzümden sana kızdı değil mi? A1 Zero-shot:؟ ࠽ ᅋᄪـڿ ۑـໞ ૯ેܰغڶ ܇܌ أ॒ु، A2 Stabilized:؟ ࠽ ᅋᄪڶ ۑـໞ ૯ેܰغڶ ܇ؽڿ עݓن أؔ، A3 Pivot:؟ ࠽ ᅋᄪ ᇖᇇّـڿ ໞܰغڶ ܇܌ أ॒ु، Observed Pattern:Role Reassignment + Blame Shift Example 5 TR:Eski düzen mi? A1 Zero-shot:ୗٔ٨ݼ؟ A2 Stabilized: ا۹ؽݟم ا۹اٴሙᇽٔ؟ A3 Pivot:ا۹ؽݟم ا۹اٴሙᇽٔ؟ Observed Pattern:Structural Completion + Nominal Recovery Example 6 TR:Kabul ediyor musunuz iş teklifimi? A1 Zero-shot:౼اٴغ߀ߐا עካ ุ༞৹ש܊۳؟ A2 Stabilized:౼اٴغ߀ߐا ע৹ض॒ुᅋᄪ ؟ A3 Pivot:౼اٴغ۳ ע৹ض وݜسوٴ්؟ Observed Pattern:Argument Structure Reweighting 14 Figure2: Aggregate behavioral comparison across translation agents. Stabilization amplifies dialect marker frequency and reduces structural expansion, while pivot me- diation introduces compression and moderate normalization effects. The zero-shot agent (A1) exhibits the highest structural variance (σ L = 2.8761) and the lowest dialect marker activation (0.2266). This pattern suggests under-specified dialect representation: without explicit reinforcement, the model oscillates between standardized and colloquial realizations, resulting in unstable decoding trajectories. Zero-shot instability therefore reflects representational uncertainty rather than translation failure. Dialect stabilization (A2) nearly doubles dialect marker frequency (0.4988) while significantly reducing structural variance (σ L = 1.1834). Notably, this shift emerges under intentionally lightweight tuning—only a single epoch with a low learning rate—highlighting how even minimal dialect exposure can substan- tially reshape decoding behavior. Given this minimal adaptation regime, the observed change reflects lexical probability redistribution rather than structural retraining. Stabilization strengthens dialectal activation while preserving se- mantic integrity, demonstrating that low-resource dialect reinforcement primar- ily operates through probability reweighting within the existing representation space. The pivot agent (A3) introduces an intermediate English representation and produces the lowest mean length ratio (0.9354) with moderate dialect activa- tion (0.2578). This configuration imposes normalization pressure, frequently compressing emotionally loaded or colloquial constructions. Pivot mediation 15 therefore acts as a regularizing constraint that reduces instability but attenu- ates dialect intensity. Taken together, these results confirm that inter-agent divergence reflects con- trolled displacement in lexical weighting and structural realization. Translation pathways function as mechanisms for exploring neighboring regions of a com- mon semantic manifold. Divergence is thus interpretable signal—revealing how multilingual systems negotiate authenticity, normalization, and mediation under different decision constraints. 8 Sentence-Level Divergence Analysis While aggregate metrics capture behavioral displacement at the corpus level, sentence-level inspection reveals how these divergences manifest in concrete translation outputs. Examples were selected from sentences with the highest inter-agent divergence scores (∆ ij ) as defined in Section 5.4. These cases there- fore represent inputs where pathway-induced behavioral displacement across agents is most pronounced. Table 3 reports the translations produced by the three agents for each selected sentence. Presenting the outputs side-by-side allows direct inspection of how different pathway constraints reshape lexical choice, syntactic realization, and semantic interpretation within the same underlying multilingual model. Across examples, zero-shot outputs frequently exhibit lexical ambiguity, struc- tural under-specification, or inconsistent register realization. Dialect stabiliza- tion increases colloquial activation while improving syntactic completeness. In contrast, pivot mediation often compresses emotionally marked constructions and occasionally introduces semantic reinterpretation through intermediate En- glish normalization. These qualitative observations are consistent with the quantitative patterns reported in Section 6, confirming that inter-agent differences reflect structured behavioral variation rather than stochastic decoding noise. 9 Implications for Explainable Multi-Agent Systems The proposed framework generalizes to generative decision systems where be- havior can be interpreted through agent-level pathway variation. Instead of probing internal activations or attention maps, interpretability emerges through relational comparison. Each agent operates under distinct con- straints (direct, stabilized, pivot-mediated), and divergence between outputs be- comes an explanatory mechanism. In this context, explainability is not extracted post-hoc. It is constructed structurally through controlled variation of translation pathways. This approach is particularly valuable in low-resource settings, where ambiguity and variability are intrinsic rather than exceptional. The multi-agent perspective also highlights that multilingual models contain latent decision flexibility. By selectively reinforcing or mediating translation 16 pathways, we expose how lexical and structural weighting shifts within the same architecture. 10 Limitations Several limitations should be acknowledged. First, the study focuses exclusively on Turkish–Syrian Arabic translation. Although the decision-space framework is conceptually transferable, empirical validation across additional language pairs and dialect continua is required to confirm generalizability. Second, dialect stabilization was intentionally lightweight (one epoch, low learning rate) to isolate lexical reweighting effects. More extensive adaptation may yield different structural dynamics, which were not explored in this study. Third, the proposed behavioral metrics emphasize divergence rather than ad- equacy. While this aligns with our interpretability objective, future work may integrate performance-oriented measures to balance authenticity with transla- tion quality. Finally, lexical overlap calculations may be affected by orthographic variation in Arabic dialect writing. Although preprocessing was controlled, dialect spelling variability remains an inherent challenge in low-resource settings. 11 Conclusion This study introduced a decision-space modeling framework for low-resource dialect translation, reframing translation pathways as constrained agents oper- ating within a shared multilingual representation space. Rather than evaluating translation as a single deterministic output, we analyzed structured behavioral displacement across zero-shot, stabilized, and pivot-mediated configurations. Empirical results over 5,000 evaluation sentences demonstrated that lightweight dialect stabilization substantially amplifies dialect marker activation while re- ducing structural instability. In contrast, pivot mediation introduces measur- able normalization and compression effects. These systematic shifts confirm that translation pathways induce controlled reweighting within the translation decision space rather than stochastic variation. By modeling translation as a multi-agent system, we show that inter-agent divergence can serve as an interpretability signal. Behavioral displacement across agents reveals how multilingual models balance dialectal authenticity, structural realization, and semantic mediation under different pathway constraints. Beyond Turkish–Syrian translation, the proposed framework suggests a broader paradigm for analyzing generative systems through controlled pathway varia- tion. Future work may extend this approach to additional dialect continua and generative tasks, further exploring pathway-based modeling as a mechanism for interpretable multilingual AI. More broadly, controlled pathway variation may serve as a practical strategy for explainable multi-agent analysis in generative language systems. 17 Acknowledgements This work was supported by the Luxembourg National Research Fund (FNR) through the project “The Epistemology of AI Systems” (C22/SC/17111440/EAI). The authors also thank NIS Studios for providing dialogue data used in the ex- periments of this study. References [1] Abbas, M., Lichouri, M., Freihat, A.A.: St madar 2019 shared task: Arabic fine-grained dialect identification. In: Proceedings of the Fourth Arabic Natural Language Processing Workshop. p. 269–273 (2019) [2] Alkheder, H., Bouamor, H., Habash, N., Zengin, A.: Benchmarking dialec- tal arabic-turkish machine translation. In: Proceedings of Machine Trans- lation Summit XIX, Vol. 1: Research Track. p. 261–271 (2023) [3] Baniata, L.H., Park, S., Park, S.B.: A neural machine translation model for arabic dialects that utilises multitask learning (mtl). Computational intelligence and neuroscience2018(1), 7534712 (2018) [4] Belinkov, Y., Glass, J.: Analysis methods in neural language processing: A survey. Transactions of the Association for Computational Linguistics7, 49–72 (2019) [5] Bi, T., Xiong, H., He, Z., Wu, H., Wang, H.: Multi-agent learning for neural machine translation. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). p. 856– 865 (2019) [6] Bouamor, H., Habash, N., Salameh, M., Zaghouani, W., Rambow, O., Abdul-Mageed, M., Obeid, O., Shahrour, A., Khalifa, S.: The madar arabic dialect corpus and lexicon. In: Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). European Language Resources Association (ELRA), Miyazaki, Japan (2018) [7] Dabre, R., Fujita, A., Chu, C.: Exploiting multilingualism through multi- stage fine-tuning for low-resource neural machine translation. In: Proceed- ings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). p. 1410–1416 (2019) [8] Ferrando, J., Gállego, G.I., Alastruey, B., Escolano, C., Costa-jussà, M.R.: Towards opening the black box of neural machine translation: Source and target interpretations of the transformer. In: Proceedings of the 2022 Con- ference on Empirical Methods in Natural Language Processing. p. 8756– 8769 (2022) [9] Gu, J., Wang, Y., Chen, Y., Li, V.O., Cho, K.: Meta-learning for low- resource neural machine translation. In: Proceedings of the 2018 conference on empirical methods in natural language processing. p. 3622–3631 (2018) 18 [10] Kim, Y., Gao, Y., Ney, H.: Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. p. 1246– 1257 (2019) [11] Kim, Y., Petrov, P., Petrushkov, P., Khadivi, S., Ney, H.: Pivot-based transfer learning for neural machine translation between non-english lan- guages. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). p. 866–876 (2019) [12] Lai, H., Toral, A., Nissim, M.: Multidimensional evaluation for text style transfer using chatgpt. arXiv preprint arXiv:2304.13462 (2023) [13] Liao, B., Gao, Y., Ney, H.: Multi-agent mutual learning at sentence-level and token-level for neural machine translation. In: Findings of the Associ- ation for Computational Linguistics: EMNLP 2020. p. 1715–1724 (2020) [14] Luo, G., Yang, Y., Yuan, Y., Chen, Z., Ainiwaer, A.: Hierarchical trans- fer learning architecture for low-resource neural machine translation. IEEE Access7, 154157–154166 (2019) [15] Madsen, A., Meade, N., Adlakha, V., Reddy, S.: Evaluating the faithfulness of importance measures in nlp by recursively masking allegedly important tokens and retraining. In: Findings of the Association for Computational Linguistics: EMNLP 2022. p. 1731–1751 (2022) [16] Mohammadshahi, A., Vamvas, J., Sennrich, R.: Investigating multi-pivot ensembling with massively multilingual machine translation models. In: Proceedings of the Fifth Workshop on Insights from Negative Results in NLP. p. 169–180 (2024) [17] Neubig, G., Hu, J.: Rapid adaptation of neural machine translation to new languages. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. p. 875–880 (2018) [18] Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for auto- matic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL). p. 311– 318 (2002) [19] Rei, R., Stewart, C., Farinha, A.C., Lavie, A.: Comet: A neural framework for mt evaluation. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). p. 2685–2702 (2020) [20] Salloum, W., Habash, N.: Dialectal arabic to english machine translation: Pivoting through modern standard arabic. In: Proceedings of the 2013 Con- ference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies. p. 348–358 (2013) [21] Shapiro, P., Duh, K.: Comparing pipelined and integrated approaches to dialectal arabic neural machine translation. In: Proceedings of the Sixth Workshop on NLP for Similar Languages, Varieties and Dialects. p. 214– 222 (2019) [22] Sibaee, S., Nacar, O., Al-Habashi, Y., Ammar, A., Boulila, W.: Shami-mt: a syrian arabic dialect to modern standard arabic bidirectional machine translation system. arXiv preprint arXiv:2508.02268 (2025) 19 [23] Talafha, B., Farhan, W., Altakrouri, A., Al-Natsheh, H.: Mawdoo3 ai at madar shared task: Arabic tweet dialect identification. In: Proceedings of the Fourth Arabic Natural Language Processing Workshop. p. 239–243 (2019) [24] Voita, E., Sennrich, R., Titov, I.: Analyzing the source and target contri- butions to predictions in neural machine translation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). p. 1126–1140 (2021) [25] Wang, X., Yin, M.: Are explanations helpful? a comparative study of the effects of explanations in ai-assisted decision-making. In: Proceedings of the 26th International Conference on Intelligent User Interfaces. p. 318– 328 (2021) [26] Wu, M., Xu, J., Yuan, Y., Haffari, G., Wan, L., Luo, W., Zhang, K.: (perhaps) beyond human translation: Harnessing multi-agent collaboration for translating ultra-long literary texts. Transactions of the Association for Computational Linguistics13, 901–922 (2025)