This paper presents an unsupervised quality estimation metric for machine translation that combines both neural and LLM models to identify and classify error spans for machine translations. As a submission to the WMT2026 Automated Evaluation Shared Task 1, this system builds on XCOMET-XL by applying offsets to the logits produced by the neural metric, post-hoc, to address the long-tailed probability problem, and uses an LLM as a second-stage to refine these predictions. Results show that XCOMET-XL with logits offset alone shows a boost in recall and character-level F1 scores for error span classifications across multiple language pairs, surpassing state-of-the-art baseline (i.e XCOMET-XL) and LLM-based approaches. This system also comprises a second-stage where a reasoning LLM-judge is employed to audit the error spans predicted by the logits-adjusted XCOMET metric, whose goal is to reduce spurious error classifications and improve overall precision. Results show that using LLM-judge to refine error spans provides no measurable improvements over predictions of the logits-adjusted XCOMET metric.

We present a proof-of-concept system for simultaneous speech translation based on dynamic attention masking. Our approach builds on SeamlessM4T by injecting lightweight per-layer schedulers into the conformer-encoder, training each scheduler to predict the number of future frames needed for translation. The schedulers are trained jointly with LoRA adapters across three language directions: English to German, Italian, and Chinese. At inference time, we evaluate our system using sliding window retranslation inference regime (Sen et al., 2022), and an adapted version of StreamAtt (Papi et al., 2024) that replaces the fixed cutoff with a content-aware threshold derived from the learnt representations from the scheduler outputs.

This system paper describes the development of a Neural Machine Translation system that is adapted to the Information Technology (IT) domain, and is able to translate specialized IT-related terminologies. Despite the popularity of incorporating terminology constraints at training time to develop terminology-aware Neural Machine Translation engines, one of the main issues is: In the absence of terminology references for training, and with the proliferation of source-target alignments, how does one select word alignments as pseudo-terminology constraints? The system in this work uses the encoder’s final hidden states as proxies for terminologies, and selects word alignments with the highest norm as pseudo-terminology constraints for inline annotation at run-time. It compares this context-based approach against a conventional statistical approach, where terminology-constraints are selected based on a low-frequency threshold. The systems were evaluated for general translation quality and Terminology Success Rates, with results that validate the effectiveness of the contextual approach.

Biomedical articles are often inaccessible to non-experts due to their technical complexity. To improve readability and factuality of lay summaries, we built on an extract-then-summarize framework by experimenting with novel extractive summarization strategies and employing Low Rank Adaptation (LoRA) fine-tuning of Meta-Llama-3-8B-Instruct on data selected by these strategies. We also explored counterfactual data augmentation and post-processing definition insertion to further enhance factual grounding and accessibility. Our best performing system treats the article’s title and keywords (i.e. search terms) as a single semantic centroid and ranks sentences by their semantic similarity to this centroid. This constrained selection of data serves as input for fine-tuning, achieving marked improvements in readability and factuality of downstream abstractive summaries while maintaining relevance. Our approach highlights the importance of quality data curation for biomedicallay summarization, resulting in 4th best overall performance and 2nd best Readability performance for the BioLaySumm 2025 Shared Task at BioNLP 2025.