We present MAGADIR’s submissions to the WMT 2026 Low-Resource Arabic–Asian Language Translation Shared Task for English–Arabic translation. Our systems combine a small official parallel corpus, parameter-efficient adaptation of open-weight multilingual models, and segment-level, reference-free quality-estimation re-ranking over heterogeneous cross-system candidate pools. For Arabic-to-English translation, the pool includes supervised LoRA-adapted NLLB and Hunyuan-family models alongside zero-shot Hunyuan/HY-MT and Gemma models. For English-to-Arabic, we select among zero-shot Hunyuan/HY-MT, Gemma, Aya, and NLLB candidates. The Arabic-to-English primary system uses MetricX-24-QE together with a Wikidata-derived entity-recall floor that vetoes candidates omitting recognized proper names before quality estimation determines the final output; the English-to-Arabic primary uses COMETKiwi without an entity constraint. On the official Challenge Test, the Arabic-to-English primary achieves 27.45 BLEU, 56.61 chrF, and 81.94 COMET, while the English-to-Arabic primary obtains 14.59 BLEU, 51.48 chrF, and 83.88 COMET. Relative to our own single-model contrastives, QE-guided cross-system selection improves COMET by up to 1.94 points for Arabic-to-English and 0.81 points for English-to-Arabic, although some contrastive systems retain stronger lexical-overlap scores. This exposes a persistent trade-off between neural adequacy objectives and reference-based surface metrics. A post-hoc matched control separates the Arabic-to-English entity floor from the otherwise confounded change in QE utility. The floor accounts for 83% of the entity-recall gain while changing only 7% of outputs and imposing effectively no COMETKiwi cost (+0.0002); the switch from COMETKiwi to MetricX rewrites 74% of outputs and accounts for the full −0.0085 COMETKiwi difference. Our audit also identifies an important limitation: on approximately 65% of entity-bearing segments, no candidate preserves every recognized entity in canonical form, leaving the veto with no feasible alternative. These findings motivate routine reference-free entity auditing, more diverse candidate pools, expanded gazetteer coverage, and transliteration-aware entity matching for QE-selected translation systems.

This work investigates the effectiveness of masked language models (MLMs) and autoregressive language models (LLMs) with fewer than one billion parameters in the detection of implicit hate speech through fine-grained span identification. The evaluation spans zero-shot, few-shot, and full supervision settings across two core benchmarks—SBIC and IHC—and an auxiliary testbed, OffensiveLang.RoBERTa-Large-355M emerges as the strongest zero-shot model, achieving the highest F1 scores of 75.8 (SBIC) and 72.5 (IHC), outperforming larger models like LLaMA 3.2-1B. ModernBERT-125M closely matches this performance with scores of 75.1 and 72.2, demonstrating the advantage of architectural efficiency. Among instruction-tuned models, SmolLM2-135M Instruct and LLaMA 3.2 1B Instruct consistently outperform their non-instructed counterparts, with up to +2.3 F1 gain on SBIC and +1.7 on IHC. Interestingly, the larger SmolLM2-360M Instruct does not outperform the 135M variant, highlighting that model scale does not always correlate with performance in implicit hate detection tasks.Few-shot fine-tuning with SmolLM2-135M Instruct achieves F1 scores of 68.2 (SBIC) and 64.0 (IHC), trailing full-data fine-tuning by only 1.6 and 2.0 points, respectively, with accuracy drops under 0.5 points. This illustrates the promise of compact, instruction-aligned models in data-scarce settings, particularly when optimized with Low-Rank Adaptation (LoRA).Topic-guided error analysis using Latent Dirichlet Allocation (LDA) reveals recurring model failures in ideologically charged or euphemistic discourse. Misclassifications often involve neutral references to identity, politics, or advocacy language, underscoring current limitations in discourse-level inference and sociopragmatic understanding.

Interaction theories suggest that the emergence of mutual understanding between speakers in natural conversations depends on the construction of a shared knowledge base (common ground), but the details of which information and the circumstances under which it is memorized are not explained by any model. Previous works have looked at metrics derived from Information Theory to quantify the dynamics of information exchanged between participants, but do not provide an efficient way to locate information that will enter the common ground. We propose a new method based on the segmentation of a conversation into themes followed by their summarization. We then obtain the location of information transfers by computing the distance between the theme summary and the different utterances produced by a speaker. We evaluate two Large Language Models (LLMs) on this pipeline, on the French conversational corpus Paco-Cheese. More generally, we explore how the recent developments in the field of LLMs provide us with the means to implement these new methods and more generally support research into questions that usually heavily relies on human annotators.

Les théories de l’interaction suggèrent que l’émergence d’une compréhension mutuelle entre les locuteurs en conversation naturelle dépend de la construction d’une base de connaissances partagée (common ground), mais n’explicitent ni le choix ni les circonstances de la mémorisation de ces informations.Des travaux antérieurs utilisant les métriques dérivées de la théorie de l’information pour analyser la dynamique d’échange d’information ne fournissent pas de moyen efficace de localiser les informations qui entreront dans le common ground. Nous proposons une nouvelle méthode basée sur la segmentation automatique d’une conversation en thèmes qui sont ensuite résumés. L’emplacement des transferts d’informations est finalement obtenu en calculant la distance entre le résumé du thème et les différents énoncés produits par un locuteur. Nous évaluons deux grands modèles de langue (LLMs) sur cette méthode, sur le corpus conversationnel français Paco-Cheese. Plus généralement, nous étudions la façon dont les derniers développement dans le champ des LLMs permettent l’étude de questions s’appuyant normalement fortement sur le jugement d’annotateurs humains.