Tag
The paper proposes query-aware routing for multilingual encoders, combining a query-only low-rank adapter (SampoTron) trained against frozen Nemotron-3-Embed-1B document embeddings with deterministic language routing. This yields a 20.9% relative nDCG@10 gain across six English, Finnish, and Swedish retrieval directions while preserving same-language performance and the existing document index.
Cactus Compute releases Whistle, a 16.9 MB open speech recognition model running dependency-free on CPU across mobile, wearables, robotics, automotive and microcontrollers. It transcribes seven languages with word timestamps and speech embeddings, reaching first token in 11 ms, and shares Needle's C++ engine and quantisation.
Aleph Alpha released Kolibri-1, an open-weight mixture-of-experts reasoning model with 78B total parameters (3.46B active) supporting up to 1M-token context, German and English focus, reasoning mode, and tool calling, released under Apache 2.0.
该论文研究了参考答案（gold reference）的选择对多语言基准评测分数的影响，发现交换参考标注可使F1分数移动数个点，并揭示了一个未记录的标注聚合默认设置导致参考答案与被评输出部分重合的问题，提出应将参考敏感性区间与分数一同报告。
该论文研究了在单语检索增强生成（RAG）中推理语言与查询/检索文档语言对齐的作用，构建了一个基于虚构桌游世界《The DarkEye》的德语 RAG 问答测试集，发现对齐推理语言能提升表现，但仅能达到模型原生英语推理的水平。
HyperBrowseComp introduces a multilingual and multimodal browsing benchmark with 423 manually authored, human-validated questions across 13 languages, designed to stress web-browsing agents through obscure evidence discovery and multi-step clue chains across videos, scanned documents, images, and maps. The authors evaluate several models under a common agent protocol and include a human baseline for comparison.
Microsoft AI 发布首个流式转录模型 MAI-Transcribe-2-Streaming（在 Artificial Analysis 上排名准确度第一，60 种语言、100ms 部分结果延迟），同时推出 MAI-Voice-2.1 与 MAI-Voice-2.1-Flash 语音合成模型，面向低延迟对话式语音智能体场景。
该论文测量了2026年主流LLM分词器对法语及法国地方语言的token溢价（法语比英语多31%–58%，地方语言高达1.6–3.3倍），并提出法语优化的字节级BPE原型Baracoda FR v1.2，在法语上比Tekken少用11.5%的token。
Bilibili's Index LLM team releases Index-Translate and companion models (Index-NativeLong, Index-Homura, Index-Echo) supporting 150 text languages, document translation, multilingual subtitles, and voice-preserving dubbing, released under Apache-2.0 with open weights in 2B, 9B, and 35B-A3B sizes.
UTTER IN is a voice-powered productivity app that turns a single voice note containing multiple requests in English or Arabic dialects into separate reminder cards, calendar events, and saved places, which the user must approve before syncing to Google, Outlook, or Apple Calendar.
BiliBili Index team announced Index-Translate, a family of multilingual translation models (2B/9B/35A3B) built on Qwen3.5 supporting 150 languages, alongside speech-to-text, speech-to-speech, syllable-constrained, and long-document translation variants released under Apache 2.0.
An arXiv paper audits TypeSafe's JEV decision-only model using the Values Survey Module 2013 across Saudi and American personas in English and Arabic, finding answers highly repeatable but persona- and language-dependent, reproducing 87% of the human Saudi–US cultural gap in English but only 62% in Arabic.
ArXiv preprint examining how Gemma 4 31B internally represents nine African languages and whether country-specific steering directions can bias English generation, finding regional alignment patterns among African languages and measurable steering effects for Nigeria, Ghana, Kenya, and South Africa.
This paper evaluates how reliably sentiment concept directions extracted from language model representations reproduce across splits, comparing English, Hausa, and Yoruba over four language models. It finds consistent language rank order in direction agreement (English > Hausa > Yoruba) and shows that high probe classification accuracy does not imply directional consistency, though it does not establish tokenizer fertility as the cause.
The paper introduces tau-Multilingual, a benchmark extending tau-Voice to Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin for evaluating voice agents across task completion, interaction quality, and language generation. Results show Spanish, Portuguese, and Hindi perform close to English, while Korean and Mandarin drop significantly, and the authors release language packs, validated judges, and tools for community-built multilingual voice-agent evaluation.
Liquid AI announced d1, their first decision model, which outperforms Jev on Hugging Face's Decision Index with advantages in multilingual evaluations, robustness, and handling longer inputs.
Hugging Face introduces the Open TTS Leaderboard, an objective, scalable evaluation framework for multilingual TTS and voice cloning using ASR-based intelligibility (WER/CER), speed (RTFx/TTFA), and WavLM speaker-similarity metrics, reducing model evaluation time from weeks to hours.
Cohere released Embed 5, its most powerful embedding family yet, with pro and fast variants sharing an embedding space, multimodal support, 100+ language coverage, 128k context, and matryoshka embeddings across multiple dimensions.
Free, open-source AI engineering course with 523 lessons, now available as EPUB/PDF books and in eight languages, focusing on building algorithms from scratch.
The paper explores using code-switched text to align representations in small language models under data constraints, demonstrating that a curriculum involving code-switching improves multilingual performance.