Adapting large language models to an individual author’s style from a few examples is challenging, and scientific writing sharpens the difficulty: formal conventions leave little surface variation, and authors write about their own topics, so extracted “style” easily entangles with content. We study style-conditioned abstract generation from a few example abstracts per author and propose three methods: (1) contrastive activation steering, (2) a network that predicts steering vectors, and (3) a hypernetwork that predicts LoRA adapters. We find a consistent trade-off between style imitation and output quality: fine-tuning buys most of the available style signal but forfeits fluency, while the hypernetwork achieves the best trade-off on both seen and unseen authors. Our steering operates at author level, contrasting an author’s abstracts against style-neutral generations for the same content. This holds topic fixed, removes the need for a predefined style inventory, and outperforms inventory-based steering. Moreover, our analyses demonstrate that manually extracted and predicted steering vectors are near-orthogonal yet score comparably, indicating that style conditioning here can admit at least two unrelated directions rather than requiring one particular axis.

We present the findings of the WMT26 Multilingual Instruction Shared Task (MIST), evaluating small (=10B parameter) open-weight language models across 24 languages on three sub-tasks: context-based QA, cross-lingual summarization, and open-ended generation. Submissions from eight teams explore fine-tuning, knowledge distillation, and prompt engineering. Our findings demonstrate that small models are not yet polyglots: while leading systems perform well on high-resource languages, performance drops sharply on low-resource varieties, and comprehension degrades significantly when transferring across non-English language pairs. In open-ended generation, human evaluation and automatic rule-based verification exhibit strong overall rank correlation (0.79) but diverge locally at the top: human evaluators favor distilled pipelines with superior fluency and cross-lingual equity, whereas automatic verifiers reward concise prompt-engineered systems. Furthermore, multi-stage pipelines suffer from prompt-language inertia on cross-lingual instructions, and summarization exhibits severe English leakage unless penalized by language gating. Finally, while baseline fluency is largely achieved by modern small models, multi-constraint instruction following and unanswerable question refusal remain the primary performance bottlenecks. We release all test sets, system outputs, human judgments, and evaluation code under a permissive license.

This paper reports on the outcomes of the shared tasks organized as part of the 23rd International Workshop on Spoken Language Translation (IWSLT). The workshop covered ten major challenges in spoken language translation, including speech-to-text translation for both high-resource and low-resource language pairs, customized speech translation, speech generation, instruction-following speech processing, and the evaluation of speech translation systems. The shared tasks received strong participation, with more than 30 teams submitting runs. This year’s edition broadened the range of tasks, placing particular emphasis on speech generation and evaluation metrics.

We present Diet-KIT, a system for the IWSLT speech translation compression task under a strict 4 GB on-disk storage constraint, starting from the 16 GB Qwen2-Audio-7B base model. Compression is achieved with a sequential pipeline based on Half-Quadratic Quantization (HQQ). Based on systematic ablations, we find that 4-bit quantization preserves translation quality well, whereas 3-bit quantization induces a sharp performance cliff, precluding aggressive compression across the whole model. We further show that the embedding table tolerates 2-bit quantization with negligible loss, while the LM head requires higher precision. To satisfy the storage constraint, we propose a sensitivity-guided layer selection method that identifies MLP sublayers tolerant to 3-bit compression via a per-layer sensitivity analysis, which consistently outperforms manual and random layer selection. Finally, AWQ calibration is applied as a data-driven refinement stage. The final system achieves 3.98 GB on disk with COMET scores of 74.4 on en→de and 77.1 on en→zh, compared to 75.6 and 79.5 for the uncompressed fine-tuned model.

Speech translation models are increasingly capable of preserving speech-specific information (e.g., speaker gender, prosody, and emphasis), yet evaluation metrics remain blind to such phenomena. We meta-evaluate both text- and speech-based quality estimation metrics on two contrastive datasets targeting gender agreement and prosody, and find that both fall short, even when given direct access to the speech signal. We then train SpeechCOMET, a family of quality estimation models with speech encoders, and evaluate a state-of-the-art SpeechLLM as a judge. Both match or exceed text-based COMET on standard quality estimation, but neither consistently assesses speech-specific phenomena. We identify three causes: (1) speech-specific features are not reliably preserved in current encoders, (2) models tend to ignore the speech source signal, and (3) quality estimation training data contains too few relevant examples. We release all models and code, and argue that progress requires dedicated speech-specific training data and models that genuinely condition on speech.

Multimodal foundation models aim to create a unified representation space that abstracts away from surface features like language syntax or modality differences. To investigate this, we study the internal representations of three recent models, analyzing the model activations from semantically equivalent sentences across languages in the text and speech modalities. Our findings reveal that: 1) Cross-modal representations converge over model layers, except in the initial layers specialized at text and speech processing. 2) Length adaptation is crucial for reducing the cross-modal gap between text and speech, although current approaches’ effectiveness is primarily limited to high-resource languages. 3) Speech exhibits larger cross-lingual differences than text. 4) For models not explicitly trained for modality-agnostic representations, the modality gap is more prominent than the language gap.

While large language models demonstrate remarkable capabilities at task-specific applications through fine-tuning, extending these benefits across diverse languages is essential for broad accessibility. However, effective cross-lingual transfer is hindered by LLM performance gaps across languages and the scarcity of fine-tuning data in many languages. Through analysis of LLM internal representations from over 1,000+ language pairs, we discover that middle layers exhibit the strongest potential for cross-lingual alignment. Building on this finding, we propose a middle-layer alignment objective integrated into task-specific training. Our experiments on slot filling, machine translation, and structured text generation show consistent improvements in cross-lingual transfer, especially to lower-resource languages. The method is robust to the choice of alignment languages and generalizes to languages unseen during alignment. Furthermore, we show that separately trained alignment modules can be merged with existing task-specific modules, improving cross-lingual capabilities without full re-training. The code is provided in the supplementary materials.

Direct speech translation (ST) models often struggle with rare words. Incorrect translation of these words can have severe consequences, impacting translation quality and user trust. While rare word translation is inherently challenging for neural models due to sparse learning signals, real-world scenarios often allow access to translations of past recordings on similar topics. To leverage these valuable resources, we propose a retrieval-and-demonstration approach to enhance rare word translation accuracy in direct ST models. First, we adapt existing ST models to incorporate retrieved examples for rare word translation, which allows the model to benefit from prepended examples, similar to in-context learning. We then develop a cross-modal (speech-to-speech, speech-to-text, text-to-text) retriever to locate suitable examples. We demonstrate that standard ST models can be effectively adapted to leverage examples for rare word translation, improving rare word translation accuracy over the baseline by 17.6% with gold examples and 8.5% with retrieved examples. Moreover, our speech-to-speech retrieval approach outperforms other modalities and exhibits higher robustness to unseen speakers. Our code is publicly available.

With the rapid development of Large Language Models (LLMs), it is crucial to have benchmarks which can evaluate the ability of LLMs on different domains. One common use of LLMs is performing tasks on scientific topics, such as writing algorithms, querying databases or giving mathematical proofs. Inspired by the way university students are evaluated on such tasks, in this paper, we propose SciEx - a benchmark consisting of university computer science exam questions, to evaluate LLMs’ ability on solving scientific tasks. SciEx is (1) multilingual, containing both English and German exams, and (2) multi-modal, containing questions that involve images, and (3) contains various types of freeform questions with different difficulty levels, due to the nature of university exams. We evaluate the performance of various state-of-the-art LLMs on our new benchmark. Since SciEx questions are freeform, it is not straightforward to evaluate LLM performance. Therefore, we provide human expert grading of the LLM outputs on SciEx. We show that the free-form exams in SciEx remain challenging for the current LLMs, where the best LLM only achieves 59.4% exam grade on average. We also provide detailed comparisons between LLM performance and student performance on SciEx. To enable future evaluation of new LLMs, we propose using LLM-as-a-judge to grade the LLM answers on SciEx. Our experiments show that, although they do not perform perfectly on solving the exams, LLMs are decent as graders, achieving 0.948 Pearson correlation with expert grading.

Diffusion models have recently shown great potential on many generative tasks.In this work, we explore diffusion models for machine translation (MT).We adapt two prominent diffusion-based text generation models, Diffusion-LM and DiffuSeq, to perform machine translation.As the diffusion models generate non-autoregressively (NAR),we draw parallels to NAR machine translation models.With a comparison to conventional Transformer-based translation models, as well as to the Levenshtein Transformer,an established NAR MT model,we show that the multimodality problem that limits NAR machine translation performance is also a challenge to diffusion models.We demonstrate that knowledge distillation from an autoregressive model improves the performance of diffusion-based MT.A thorough analysis on the translation quality of inputs of different lengths shows that the diffusion models struggle more on long-range dependencies than other models.

Customizing machine translation models to comply with desired attributes (e.g., formality or grammatical gender) is a well-studied topic. However, most current approaches rely on (semi-)supervised data with attribute annotations. This data scarcity bottlenecks democratizing such customization possibilities to a wider range of languages, particularly lower-resource ones. This gap is out of sync with recent progress in pretrained massively multilingual translation models. In response, we transfer the attribute controlling capabilities to languages without attribute-annotated data with an NLLB-200 model as a foundation. Inspired by techniques from controllable generation, we employ a gradient-based inference-time controller to steer the pretrained model. The controller transfers well to zero-shot conditions, as it is operates on pretrained multilingual representations and is attribute- rather than language-specific. With a comprehensive comparison to finetuning-based control, we demonstrate that, despite finetuning’s clear dominance in supervised settings, the gap to inference-time control closes when moving to zero-shot conditions, especially with new and distant target languages. The latter also shows stronger domain robustness. We further show that our inference-time control complements finetuning. Moreover, a human evaluation on a real low-resource language, Bengali, confirms our findings. Our code is in the supplementary material.

With the number of scientific papers published every year growing and current large language models (LLMs) showing state-of-the-art performance on natural language processing (NLP) tasks, we ask the question if LLMs could be utilized to answer questions on scientific papers.We investigate how well state-of-the-art large language models (LLMs) can answer questions on scientific paper by experimenting with long-context versions of the LLaMA 2 model and evaluating and training on the Qasper dataset.We analyze how well the LLMs handle longer papers and questions that can only be answered by accessing information from far out paragraphs. During our experiments, we see that the performance of these LLMs drops with growing length and position of relevant information.We employ different measures from simple prompts to chain-of-thought prompts and zero-shot usage to fine-tuning with QLoRA.While we still observe a performance loss with increased context length, our measures reduce the effects of this flaw, and we can achieve F₁ scores similar to bigger models like GPT-4.