multilingual

该论文通过六语言析因设计，独立考察问题语言、回答语言与验证指令语言对多语言 P(True) 自评估的影响，发现三者分别影响准确率、区分度与校准，且未经验证的评判者会严重扭曲模型表现差距。

X-Dubbing 提出一种联合时间感知语音转文本翻译与视频感知文本转语音的方法，通过语言无关的时间预算实现与源视频同步的跨语言配音，无需配对配音数据即可超越现有基线。

This report introduces Kolibri, an English-German mixture-of-experts transformer with 78.1B total and 3.46B active parameters, released as open weights under Apache 2.0 for sovereign, German-focused deployment in regulated domains. Trained on 24T tokens with over 20% German data and post-trained with SFT plus RL on 1.2M curated tasks, Kolibri lies on the Pareto frontier of quality and serving cost among evaluated open models.

The paper extends PolyRefuse to three additional low-resourced languages (Belarusian, Bashkir, Tajik) on Qwen2.5-14B-Instruct, finding that refusal direction transfer is graded rather than universal: directions from safety-aligned languages transfer broadly, while those from strongly misaligned languages fail to reliably modulate refusal behavior.

KittenML 开源发布了 Kitten-TTS 2，这是全球首个 1.58-bit（三值化权重）TTS 模型，1.7B 参数，可在纯 CPU 上实时推理，支持少于 5 秒参考音频的零样本语音克隆、20+ 语言多语种支持以及情感与强调控制。

IBM released Granite Speech 5.0 2B, a compact multilingual speech-language model for ASR and speech translation across six languages, with improved noise robustness, language-ID prompting, and up to 2000-keyword biasing over its predecessor.

inclusionAI released Ming-3.0-TTS-0.5B, a dense 0.5B multilingual text-to-speech model built on Qwen2.5-0.5B supporting zero-shot voice cloning, voice design, paralinguistic inline tags, and free-form style control across 7 languages plus Chinese dialects.

IBM 发布 Granite Speech 5.0 系列更新，包括新的 2B 语音语言模型以及改进的 470m-turboctc 和 4.1-2b-plus 版本，在转录、翻译、抗噪和说话人归属方面更准确快速，采用 Apache 2.0 许可。

该论文提出一个翻译绕过威胁模型，在79种语言上评估了十个主流大模型，发现即使提示在英文下被安全拒绝，翻译成低资源语言后仍有21.9%至44.4%的绕过率，表明英文安全对齐无法可靠迁移到其他语言。

The paper introduces LoWAgent, a low-resource multilingual benchmark evaluating web agents across English, Hindi, Bengali, Urdu, and Amharic on six localized websites with 3,000 tasks, finding notable performance drops for low-resource languages—especially open-source models—and showing that post-training on benchmark traces improves agentic performance.

This paper introduces SITO, a synthetic independent benchmark of 2,332 contrastive cases in English, Ukrainian, and Surzhyk, used to audit the Tuteliq Detect commercial text-safety API for missed harm and false positives across six harm categories, finding notably weaker recall for third-party bystander reports of ongoing harm.

Ant Group’s inclusionAI has open-sourced the Ming-3.0-tts speech synthesis model, which ranked first overall in pronunciation robustness with a score of 93.2% on Artificial Analysis’s October leaderboard and first among open-source models for voice controllability. It supports voice cloning, control through natural-language instructions, 14 paralinguistic tags, and speech synthesis across multiple dialects and languages, including Sichuanese and Cantonese.

ElevenLabs 推出 Eleven Realtime，一套端到端优化的 Speech-to-Speech 语音模型（v1 Mini 与 v1 Max），支持实时语音对话，并提供与 OpenAI Realtime API 兼容的接口，可在 ElevenAgents 中快速构建语音助手。

Mistral 发布 Voxtral 3 语音模型家族与语音代理 SDK，支持 50+ 语言的实时转录、情感自适应语音合成及多步工具调用，可通过欧盟托管 API 或自部署使用以保障数据主权。

Qwen-Audio-3.1-ASR 技术报告介绍了一系列音频语言模型，在语音识别、说话人分离转录和通用音频理解方面较 Qwen-Audio-3.0-ASR 有显著提升，支持 30 种语言和 16 种汉语方言，并提供流式与非流式等多种部署变体。

The paper proposes the Reliance Hypothesis: LLM chain-of-thought faithfulness increases in lower-resource languages on knowledge-intensive tasks, because models rely more on reasoning when they cannot recall the answer. The effect is replicated across four models and several faithfulness measures, with model confidence partially explaining resource-level effects.

该论文对大语言模型的位置偏差进行了全面研究，覆盖从十亿到万亿参数规模的模型及五种语言。研究发现模型规模增大、思维链提示以及新兴的思考模型能显著缓解甚至几乎消除位置偏差。

This paper systematically investigates how well LLMs detect culture-specific stereotypes across prompting language, instruction tuning language, and fine-tuning dataset size for US, Dutch, and Chinese contexts, finding that LLMs overpredict stereotypes, rarely account for cultural context, and that non-English prompts and instruction tuning can improve results.

这篇论文审计了三类非生成式决策模型（编码器打分器、解码器打分器和闭源 API）在孟加拉语习语上的表现，发现当选项列表中包含字面释义时，模型会强烈偏向字面含义，且这种"字面吸引"在孟加拉语测试中比英语高 1.9 到 6.4 倍，同时校准性全面下降。

作者测试了 Embedding Gemma 2（Q8 约 330MB、768 维）在视觉小说截图上的文搜图能力，发现其对跨语言文本含义有一定理解且表现可用，但文本内容往往主导余弦相似度、视觉特征影响较弱，且需要精细的措辞。