Large Language Models (LLMs) demonstrate strong reasoning capabilities for many tasks, often by explicitly decomposing the task via Chain-of-Thought (CoT) reasoning. Recent work on LLM-based translation designs hand-crafted prompts to decompose translation, or trains models to incorporate intermediate steps. _Translating Step-by-step_ (Briakou et al., 2024), for instance, introduces a multi-step prompt with decomposition and refinement of translation with LLMs, which achieved state-of-the-art results on WMT24 test data. In this work, we scrutinise this strategy’s effectiveness. Empirically, we find no clear evidence that performance gains stem from explicitly decomposing the translation process via CoT, at least for the models on test; and we show prompting LLMs to “translate again” and self-refine yields even better results than human-like step-by-step prompting. While the decomposition influences translation behaviour, faithfulness to the decomposition has both positive and negative effects on translation. Our analysis therefore suggests a divergence between the optimal translation strategies for humans and LLMs.

The increasing use of Artificial Intelligence(AI) technologies, such as Large LanguageModels (LLMs) has led to nontrivial improvementsin various tasks, including accurate authorshipidentification of documents. However,while LLMs improve such defense techniques,they also simultaneously provide a vehicle formalicious actors to launch new attack vectors.To combat this security risk, we evaluate theadversarial robustness of authorship models(specifically an authorship verification model)to potent LLM-based attacks. These attacksinclude untargeted methods - authorship obfuscationand targeted methods - authorshipimpersonation. For both attacks, the objectiveis to mask or mimic the writing style of an authorwhile preserving the original texts’ semantics,respectively. Thus, we perturb an accurateauthorship verification model, and achievemaximum attack success rates of 92% and 78%for both obfuscation and impersonation attacks,respectively.

This paper presents UvA-MT’s submission to the WMT 2025 shared task on general machine translation, competing in the unconstrained track across all 16 translation directions. Unusually, this year we use only WMT25’s blind test set (source sentences only) to generate synthetic data for LLM training, and translations are produced using pure beam search for submission. Overall, our approach can be seen as a special variant of data distillation, motivated by two key considerations: (1) perfect domain alignment, where the training and test domains are distributionally identical; and (2) the strong teacher model, GPT-4o-mini, offers high-quality outputs as both a reliable reference and a fallback in case of mere memorization.Interestingly, the outputs of the resulting model, trained on Gemma3-12B using Best-of-N (BoN) outputs from GPT-4o-mini, outperform both original BoN outputs from GPT-4o-mini and Gemma3-12B in some high-resource languages across various metrics. We attribute this to a successful model ensemble, where the student model (Gemma3-12B) retains the strengths of the teacher (GPT-4o-mini) while implicitly avoiding their flaws.

Current machine translation (MT) systems perform well in the domains on which they were trained, but adaptation to unseen domains remains a challenge. Rather than fine-tuning on domain data or modifying the architecture for training, an alternative approach exploits large language models (LLMs), which are performant across NLP tasks especially when presented with in-context examples. We focus on adapting a pre-trained LLM to a domain at inference through in-context example selection. For MT, examples are usually randomly selected from a development set. Some more recent methods though select using the more intuitive basis of test source similarity. We employ topic models to select examples based on abstract semantic relationships below the level of a domain. We test the relevance of these statistical models and use them to select informative examples even for out-of-domain inputs, experimenting on 7 diverse domains and 11 language pairs of differing resourcedness. Our method outperforms baselines on challenging multilingual out-of-domain tests, though it does not match performance with strong baselines for the in-language setting. We find that adding few-shot examples and related keywords consistently improves translation quality, that example diversity must be balanced with source similarity, and that our pipeline is overly restrictive for example selection when a targeted development set is available.

Fine-tuning Large Language Models (FT-LLMs) with parallel data has emerged as a promising paradigm in recent machine translation research. In this paper, we explore the effectiveness of FT-LLMs and compare them to traditional encoder-decoder Neural Machine Translation (NMT) systems under the WMT24 general MT shared task for English to Chinese direction. We implement several techniques, including Quality Estimation (QE) data filtering, supervised fine-tuning, and post-editing that integrate NMT systems with LLMs. We demonstrate that fine-tuning LLaMA2 on a high-quality but relatively small bitext dataset (100K) yields COMET results comparable to much smaller encoder-decoder NMT systems trained on over 22 million bitexts. However, this approach largely underperforms on surface-level metrics like BLEU and ChrF. We further control the data quality using the COMET-based quality estimation method. Our experiments show that 1) filtering low COMET scores largely improves encoder-decoder systems, but 2) no clear gains are observed for LLMs when further refining the fine-tuning set. Finally, we show that combining NMT systems with LLMs via post-editing generally yields the best performance for the WMT24 official test set.

We present a lightweight method for identifying currently trending terms in relation to a known prior of terms, using a weighted log-odds ratio with an informative prior. We apply this method to a dataset of posts from an English-language underground hacking forum, spanning over ten years of activity, with posts containing misspellings, orthographic variation, acronyms, and slang. Our statistical approach supports analysis of linguistic change and discussion topics over time, without a requirement to train a topic model for each time interval for analysis. We evaluate the approach by comparing the results to TF-IDF using the discounted cumulative gain metric with human annotations, finding our method outperforms TF-IDF on information retrieval.