This paper describes TransCore, the NewTranx submission to the WMT26 General Machine Translation Shared Task constrained track. Our system is built upon the Qwen3-8B large language model and adopts a two-stage supervised fine-tuning framework trained exclusively on publicly available data. Our approach focuses on high-quality multilingual data construction, including parallel corpus construction, knowledge-distilled corpus construction, and instruction-following data construction. For the parallel corpus, we construct a high-quality multilingual training corpus through multilingual data collection and automatic quality filtering. For knowledge distillation, we construct a KD corpus from publicly available monolingual resources using a large language model, followed by automatic quality filtering. To support the instruction-following evaluation introduced in WMT26, we further construct instruction data covering immutable entity preservation, document-level translation, and structured translation, which are incorporated into the first-stage supervised fine-tuning. Our system participates in four language directions of the WMT26 General Machine Translation Shared Task: English$\rightarrow$Indonesian, English$\rightarrow$Kazakh, English$\rightarrow$Thai, and Chinese$\rightarrow$Japanese.

We present the findings of the WMT26 Chinese–Southeast Asian Multilingual Machine Translation shared task. The task evaluates fourteen translation directions between Chinese and Thai, Vietnamese, Lao, Burmese, Khmer, Indonesian, and Malay. Organizers released manually aligned parallel data, domain-matched monolingual data, validation data, and a blind test set. The published leaderboard contains eleven systems and combines translation quality, measured by the equal-weight average of sacreBLEU and COMET, with a throughput bonus measured on a shared four-GPU platform. Jiutian-MT ranks first overall, while AiBabelMT achieves the highest throughput. We describe the task design, data resources, evaluation protocol, and published aggregate results, then examine how quality and throughput shape the final ranking. The results also motivate more detailed reporting by language direction and domain

In real-world Tool-Integrated Reasoning (TIR) scenarios, a major source of inefficiency is that the toolcalls create pauses between LLM requests and cause KV-cache eviction. Also, the long, unfiltered response returned by external tools inflates the KV-cache, so each decode step spends more time loading the growing cache and thus becomes steadily slower as context length increases. However, existing efficiency metrics like token counts and toolcall counts fail to capture this real computational cost. To address this, we introduce PTE (Prefill Token Equivalents), a hardware-aware TIR-efficiency metric that unifies internal reasoning and external tool-use costs while explicitly accounting for non-reusable KV-Cache and long-tool-response scenarios, thus better reflects real-world scenarios. We conduct extensive experiments across five TIR benchmarks, quantify their PTE costs, and identify four inefficiency patterns that appear in TIR. In a simulated high-concurrency industrial setting, PTE explains wall-clock latency significantly better than token-count metric. We also discover that trajectories with higher PTE costs tend to have lower reasoning correctness, indicating that simply using more tools does not improve the quality of the answer. PTE offers a new perspective on the efficiency of Tool-Integrated Reasoning. The code is available.