We present Lumen, our submission to the WMT2026 General Machine Translation Shared Task. We focus on instruction-specialized translation, where systems must produce accurate translations while satisfying fine-grained requirements on style, terminology, and structured output. Lumen follows a four-stage pipeline. First, continued pre-training and supervised fine-tuning jointly strengthen multilingual translation competence and the ability to follow translation-specific instructions, such as terminology, style, and format constraints. Second, we extend the originally offline M{2}PO preference-optimization framework into an online preference-optimization stage: the current policy samples candidate translations, an LLM judge assigns continuous direct-assessment scores, and the resulting multi-pair preferences are used to update the model. Third, on-policy distillation transfers stronger translation-instruction-following behavior from a larger teacher model to the Qwen3-14B student, improving translation-instruction adherence while retaining a 14B primary translation backbone. Finally, structure-aware inference and postprocessing preserve required formats, while selective repair addresses clear structural or decoding failures before submission. We participate in 23 translation directions. Under an absolute 1–5 LLM-judge rubric that evaluates translation-instruction following and translation quality separately, Lumen reaches an overall mean score of 4.60. Despite using a 14B translation backbone, it achieves performance comparable to leading proprietary systems in both translation quality and translation-instruction following.

Discrete diffusion language models generate text by iteratively committing masked positions, making decoding a trajectory-level process rather than a sequence of independent token decisions. Existing analyses often focus on surface-level knobs such as temperature, step budget, block size, or scheduler choice, but these knobs do not explain why the same change can induce different behaviors across models. We analyze discrete diffusion decoding through high-influence commitment events and define an observed trajectory influence measure based on downstream entropy reduction over unresolved positions. Across Dream-v0-Instruct-7B and LLaDA-8B-Instruct on reasoning, code, and question-answering tasks, we find that influence is sparse, early relative to each model’s commit frontier, and stable under tested seed and step-budget variations. We further show that criticality emerges from the interaction between commit timing and token role, explaining why decoding knobs reshape trajectories in model-dependent ways. Finally, lightweight margin-based interventions validate the analysis by correcting diagnosed failure modes without additional forward passes.

Diffusion language models generate text by denoising a masked canvas, but usually require the canvas length to be fixed before decoding. This leads to an allocation dilemma: short canvases may truncate outputs, while overlong canvases waste computation and introduce noisy continuations. Existing variable-length methods adapt the canvas through expansion or length search, but largely ignore whether such intervention is safe. We argue that canvas editing can perturb partially stabilized denoising trajectories, and propose Safe Adaptive Canvas Decoding (SACD), a training-free framework for conservative canvas control. SACD diagnoses allocation, estimates intervention risk, and chooses among expand, shrink, and keep. Experiments across reasoning, code generation, and infilling tasks show that SACD improves accuracy, latency, and token utilization across multiple DLM backbones, highlighting safe sparse intervention as a key principle for variable-length diffusion decoding.

Expanding Large Language Models(LLMs) to new languages is a costly endeavor, demanding extensive Continued Pre-Training(CPT) and data-intensive alignment. While recent data-free merging techniques attempt to bypass alignment by fusing a multilingual CPT-enhanced model with its instruct counterpart, they are plagued by a critical trade-off: mitigating parameter conflicts to preserve original abilities inevitably dilutes new language acquisition, and vice-versa. To resolve this conflict, we introduce , which upcycles a dense model into a Mixture-of-Experts(MoE) architecture, allocating different experts to different languages. Alignment ability is then transferred by grafting a MoE-expanded parameter delta(𝛥_instruct) to the CPT-enhanced base model, bypassing the complex alignment phase. Experiments demonstrate ’s superiority even against baselines with similar FLOPs or number of parameters; it improves performance on expanded languages while effectively preserving original capabilities. We further show our approach is highly applicable across different models and Post-training deltas.

Large language models (LLMs) exhibit remarkable multilingual capabilities despite English-dominated pre-training, attributed to cross-lingual mechanisms during pre-training. Existing methods for enhancing cross-lingual transfer remain constrained by parallel resources, suffering from limited linguistic and domain coverage. We propose Cross-lingual In-context Pre-training (CrossIC-PT), a simple and scalable approach that enhances cross-lingual transfer by leveraging semantically related bilingual texts via simple next-word prediction. We construct CrossIC-PT samples by interleaving semantic-related bilingual Wikipedia documents into a single context window. To access window size constraints, we implement a systematic segmentation policy to split long bilingual document pairs into chunks while adjusting the sliding window mechanism to preserve contextual coherence. We further extend data availability through a semantic retrieval framework to construct CrossIC-PT samples from web-crawled corpus. Experimental results demonstrate that CrossIC-PT improves multilingual performance on three models (Llama-3.1-8B, Qwen2.5-7B, and Qwen2.5-1.5B) across six target languages, yielding performance gains of 3.79%, 3.99%, and 1.95%, respectively, with additional improvements after data augmentation.

As Large Language Models (LLMs) become more advanced, the security risks they pose also increase. Ensuring that LLM behavior aligns with human values, particularly in mitigating jailbreak attacks with elusive and implicit intentions, has become a significant challenge. To address this issue, we propose a jailbreak defense method called Real Intentions Defense (RID), which involves two phases: soft extraction and hard deletion. In the soft extraction phase, LLMs are leveraged to extract unbiased, genuine intentions, while in the hard deletion phase, a greedy gradient-based algorithm is used to remove the least important parts of a sentence, based on the insight that words with smaller gradients have less impact on its meaning. We conduct extensive experiments on Vicuna and Llama2 models using eight state-of-the-art jailbreak attacks and six benchmark datasets. Our results show a significant reduction in both Attack Success Rate (ASR) and Harmful Score of jailbreak attacks, while maintaining overall model performance. Further analysis sheds light on the underlying mechanisms of our approach.

When consumers’ shopping needs are concentrated, they are more interested in the collection of commodities under the specific marketing theme. Therefore, mining marketing themes and their commodities collections can help customers save shopping costs and improve user clicks and purchases for recommendation system. However, the current system invites experts to write marketing themes and select the relevant commodities, which suffer from difficulty in mass production, poor timeliness and low online indicators. Therefore, we propose a automatic marketing theme and commodity construction system, which can not only generate popular marketing themes and select the relevant commodities automatically, but also improve the theme online effectiveness in the recommendation system. Specifically, we firstly utilize the pretrained language model to generate the marketing themes. And then, we utilize the theme-commodity consistency module to select the relevant commodities for the above generative theme. What’s more, we also build the indicator simulator to evaluate the effectiveness of the above generative theme. When the indicator is lower, the above selective commodities will be input into the theme-rewriter module to generate more efficient marketing themes. Finally, we utilize the human screening to control the system quality. Both the offline experiments and online A/B test demonstrate the superior performance of our proposed system compared with state-of-the-art methods.