Tracking an 8.4B-parameter model through three post-training phases on Japanese financial documents, one study finds the stages interact non-additively: math-only distillation can overwrite financial knowledge already acquired, even as procedural math skill survives. Alongside that, UvA-MT describes its unconstrained WMT 2026 submission across all 23 translation directions, using sequential sampling with refinement prompts and pairwise LLM judges, while a third paper turns to controlled Russian text simplification for A1, A2 and B1 levels with Rejection Sampling and DPO.

UvA-MT's Participation in the WMT26 General Translation Shared Task

For unconstrained MT with LLMs, sequential sampling plus a refinement prompt and round-robin pairwise judging beat pointwise best-of-N selection on WMT25 offline evaluation.

This paper presents UvA-MT's submission to the WMT 2026 General Machine Translation shared task, competing in the unconstrained track across all 23 translation directions. This year, we fully leverage the test-time methods of Large Language Models (LLMs) for machine translation. Specifically, (1) we use sequential sampling with a refinement prompt to generate a pool of translation candidates for each source segment, a strategy shown to outperform traditional parallel sampling; and (2) we employ LLMs as pairwise judges to select the best candidate via a round-robin voting mechanism, with offline evaluation on the WMT25 benchmark showing a very clear improvement over point-wise best-of-N selection. Combined, these two strategies form our WMT26 submission.

CEFR-oriented Russian Text Simplification Using Reinforcement Learning

For controlled simplification, DPO on YandexGPT reaches 87.15% CEFR target accuracy and works best at A2 and B1, but A1 remains the hardest level.

We present a study on controlled automatic simplification of Russian texts using preference-based optimization algorithms with a focus on adapting texts to the A1, A2 and B1 proficiency levels. We fine-tune YandexGPT using Rejection Sampling and Direct Preference Optimization (DPO) methods to improve control over the complexity of the generated text. We compare several training configurations that differ in dataset composition, text type and the strategy used to construct preference pairs for DPO. For automatic evaluation we train a regression model based on ruRoBERTa-large to predict continuous complexity scores corresponding to CEFR levels. We also consider semantic similarity, lexical coverage and quantitative readability metrics. Human evaluation assesses simplicity, fluency and adequacy. Our best model achieves 87.15% target level accuracy, an MAE of 0.35 CEFR levels and a target interval MAE of 0.11. Manual assessment shows that 75% of the evaluated simplifications match the target morphological and syntactic level, and fluency and adequacy scores are especially high for A2 and B1. Although A1 remains the most challenging level, the model significantly reduces the proportion of vocabulary and grammatical constructions beyond the target level while generally preserving meaning. Our model is freely available on HuggingFace.

Staged Continual Adaptation of Multimodal Foundation Models for Japanese Financial Documents

For Japanese financial document models, staged post-training does not add capabilities cleanly, so you need to choose checkpoints by task because math distillation can erase earlier domain gains.

Staged post-training of multimodal foundation models chains vision-language alignment, reasoning distillation, and domain tuning, yet how each phase trades capabilities against the others remains under-characterised. Tracking an 8.4B-parameter model from an un-fine-tuned baseline through three training phases on Japanese financial documents (evaluated on five text-input benchmarks), we find the phases interact non-additively: math-only distillation overwrites the financial knowledge acquired earlier even as procedural math skill survives, and the domain phase repairs more than it installs. A skip-phase ablation is best for domain knowledge but forfeits numerical-reasoning gains, so per-phase contributions are not decomposable and the best checkpoint is task-dependent, with implications for checkpoint retention and scheduling rather than added stages.

Papers announced by arXiv on Tuesday, 20 October 2026, covering submissions from 19 October to 20 October 2026.