This submission addresses WMT26 Category 2 for five extremely low-resource English-to-X translation directions: Bodo, Karbi, Kokborok, Nagamese, and Tagin. We develop an LLM-driven pipeline in which large language models are used both to construct bilingual lexical resources and as translation models. First, an English seed dictionary and qwen3-max are used to build word-level bilingual vocabularies, whose top candidates are injected into translation prompts as a "Helpful Vocabulary" block. On an 80/20 train-held-out split of the Tagin data, vocabulary injection is associated with a 7.09-point BLEU improvement and a 49.35-point TER reduction. Second, we compare LoRA adaptation of Qwen2.5-32B-Instruct with full-parameter fine-tuning of Hunyuan-MT-7B, with optional DPO. Hunyuan performs better in the reported Bodo and Kokborok comparisons, including an 11.87-point ChrF advantage on Bodo. Third, post-hoc repeated-ngram and length truncation, combined with per-sentence selection of the shorter of two decoding outputs, mitigates generation degeneration and improves Bodo BLEU from 19.27 to 27.18 without retraining. We submit two contrastive systems per direction: a take-shorter system with output truncation for Bodo, and augmented and baseline SFT systems for the other directions. Code and configurations are publicly available.

This submission addresses WMT26 Category 2 for five extremely low-resource English-to-X translation directions: Bodo, Karbi, Kokborok, Nagamese, and Tagin. We develop an LLM-driven pipeline in which large language models are used both to construct bilingual lexical resources and as translation models. First, an English seed dictionary and qwen3-max are used to build word-level bilingual vocabularies, whose top candidates are injected into translation prompts as a "Helpful Vocabulary" block. On an 80/20 train-held-out split of the Tagin data, vocabulary injection is associated with a 7.09-point BLEU improvement and a 49.35-point TER reduction. Second, we compare LoRA adaptation of Qwen2.5-32B-Instruct with full-parameter fine-tuning of Hunyuan-MT-7B, with optional DPO. Hunyuan performs better in the reported Bodo and Kokborok comparisons, including an 11.87-point ChrF advantage on Bodo. Third, post-hoc repeated-ngram and length truncation, combined with per-sentence selection of the shorter of two decoding outputs, mitigates generation degeneration and improves Bodo BLEU from 19.27 to 27.18 without retraining. We submit two contrastive systems per direction: a take-shorter system with output truncation for Bodo, and augmented and baseline SFT systems for the other directions. Code and configurations are publicly available.

Large language models (LLMs) have recently advanced knowledge graph question answering (KGQA), but current methods tend to rely on LLM-induced type systems with inconsistent granularity, or perform multi-hop reasoning without explicit target-type constraints. We introduce OntGQA, a type-constrained KGQA framework that reasons over a relation-centric ontology graph, where each relation is labeled with its head and tail entity types to provide a stable schema backbone. Built on this graph, OntGQA adopts a planner–judge architecture with generative backoff: a type planner proposes plausible head–tail type pairs, a judge verifies retrieved candidates and their paths, and a generator is invoked only when all candidates are rejected. By constraining both endpoints of reasoning in type space, OntGQA achieves state-of-the-art performance and produces ontology-grounded reasoning chains, with substantial Hit@1 gains (87.7%→91.5% on WebQSP and 67.6%→74.6% on CWQ).

Transformer-based models have made tremendous impacts in natural language generation. However the inference speed is a bottleneck due to large model size and intensive computing involved in auto-regressive decoding process. We develop FastSeq framework to accelerate sequence generation without accuracy loss. The proposed optimization techniques include an attention cache optimization, an efficient algorithm for detecting repeated n-grams, and an asynchronous generation pipeline with parallel I/O. These optimizations are general enough to be applicable to Transformer-based models (e.g., T5, GPT2, and UniLM). Our benchmark results on a set of widely used and diverse models demonstrate 4-9x inference speed gain. Additionally, FastSeq is easy to use with a simple one-line code change. The source code is available at https://github.com/microsoft/fastseq.