This paper describes Jiutian-MT-8B, our submission to the WMT 2026 Chinese-Southeast Asian Multilingual Machine Translation shared task. The system adapts the 8-billion-parameter Jiutian-8B-Instruct checkpoint to fourteen translation directions between Chinese and Thai, Vietnamese, Lao, Burmese, Khmer, Indonesian, and Malay. We formulate translation as a tightly controlled instruction-following task and use a four-stage full-parameter training recipe: (i) supervised multilingual adaptation on released parallel data, (ii) faithfulness first confidence-gated multi-pair preference alignment (F2MTA), (iii) direction-coupled, evidence-filtered self-training from task monolingual data, and (iv) a final clean-data calibration stage. The preference stage uses a partial order: a candidate must gain source adequacy without materially sacrificing target fluency, rather than merely win a scalarized quality score. The recipe directly updates all Jiutian-8B parameters with sharded bf16 training while retaining authentic data anchors to limit synthetic-data drift. The official test set is blind at the time of writing; numerical main results are therefore deliberately unreported. We nevertheless specify the data scope, translation interface, training protocol, and evaluation plan needed to complete the report once official scores are returned.

Recent Multimodal Large Language Models (MLLMs) have achieved remarkable advancements, yet their internal mechanisms for concurrently processing diverse modalities like text, image, and audio remain largely opaque. In this paper, we propose a methodology to convert dense MLLMs into fine-grained Mixture-of-Experts (MoE) architectures. This allows us to visually investigate their multimodal activation patterns through expert activation frequency heatmaps. Conducting comprehensive experiments on representative MLLMs, we analyze the similarities and differences in internal neuron activations when handling distinct modalities. Specifically, we examine the distribution of high-frequency activated experts, the distinct roles of high-frequency (e.g., fundamental logic) and low-frequency (e.g., domain-specific concepts) multimodal shared experts, and the prevalence and localization of modality-specific experts. Furthermore, we explore leveraging these discovered activation discrepancies to guide sparse activation and model pruning. Experimental results demonstrate that our approach substantially outperforms random expert pruning and can achieve comparable or even superior performance to the original unpruned models while utilizing significantly fewer active parameters. Our work not only sheds light on the multimodal processing mechanisms within MLLMs but also provides a practical pathway toward developing more interpretable and efficient multimodal systems.