The WMT25 Multilingual Instruction Shared Task (MIST) introduces a benchmark to evaluate large language models across 30 languages, covering five problem types, and highlights limitations in automatic evaluation while providing a standardized framework for future progress.

A new lightweight framework for prompt-space optimization improves multilingual performance in large language models by transforming translated prompts for naturalness, cultural adaptation, and difficulty enhancement, achieving substantial downstream improvements over translation-only baselines.

EAGer, a training-free method, reduces computation and improves performance by branching to multiple reasoning paths only when high-entropy tokens are present, reallocating compute budget to instances needing exploration, achieving up to 65% fewer tokens and 37% improvement in Pass@k.

Fusion-of-N (FusioN) is a new method that uses a general LLM judge to synthesize the most informative elements of each sample into a single final answer. It consistently outperforms Best-of-N (BoN) across 11 languages, 3 diverse tasks and varying model scales.

Synthetic data generation for code models faces a "verification ceiling" due to verifier limitations. Richer test suites, relaxed pass thresholds, and diverse solutions improve performance. Calibrated verification with challenging problem-solution pairs can overcome this ceiling.

NeoBabel is a multilingual image generation framework that supports 6 languages and achieves state-of-the-art performance on multilingual benchmarks while maintaining strong English capability.

We study robust scaling for open-ended generative tasks in multilingual, multi-task settings. Our findings show that sampling and selection strategies must adapt to diverse domains and languages. We propose novel strategies, yielding notable gains across languages and tasks.

This work optimizes training protocols to enhance performance on underrepresented use cases, achieving up to 14.1% relative improvement on tasks like CodeRepair.