multilingual

This academic paper examines geopolitical biases and divisions that manifest across different languages in large language models, analyzing how these models handle sensitive geopolitical topics.

This paper introduces a three-layer checklist-and-judge framework to evaluate the communicative success of interpreter agents in machine-translated conversations across semantic, pragmatic, and cultural-social dimensions, validated through extensive benchmarks.

This research paper investigates why pretraining in large language models fails to transfer knowledge across languages, identifies disjoint token spaces as a fundamental barrier, and proposes mapping languages to a shared token space to improve cross-lingual generalization.

The post discusses how companies like ElevenLabs and OpenAI have expanded beyond core AI voice technology, leading to slower voice progress, and introduces Cartesia's Multilingual Voices for businesses to maintain a consistent brand voice across languages.

Karan Goel introduces Multilingual Voices, partnering with Baklavastory to demonstrate how businesses can maintain consistent character across different languages and calls.

The post discusses how AI is making voice the default input layer for software and introduces the Multilingual Voices feature to maintain business character across languages.

The tweet discusses how vertical SaaS is evolving to become voice-first with the help of faster AI voice models, opening new opportunities for developers, and highlights multilingual capabilities from Cartesia.

VoiceCap is an AI notetaker that transcribes meetings in over 100 languages and writes summaries, action items, and decisions in the same language. It integrates with apps like Zoom, Meet, and Teams, with EU-based storage for privacy.

M-SQE proposes a multilingual skill quality estimation framework to enhance language equality in agentic skill use by evaluating candidates through intrinsic quality and task-grounded utility.

This paper benchmarks Knowledge Graph-based augmentation against Retrieval Augmented Generation for culturally specific question answering, demonstrating significant error reduction on Latin American culture data using the LatamQA dataset.

NVIDIA introduces SEA-LION-v4.8, a family of multilingual AI models built on Nemotron 3, optimized for Southeast Asian languages with significant improvements in performance on tasks like instruction following and reasoning.

NetEase Youdao AI is open-sourcing Confucius4-R2T2, a 1.7B frontier real-time streaming ASR model designed for voice agents, featuring low latency, high accuracy, and configurable decoding chunks.

The article highlights the impressive capabilities of the Gemini 3.8 live AI model, showcased by a real-time insurance claim agent that supports voice and multilingual interactions, with an open-source implementation.

This paper proposes a discrete diffusion framework for efficient one-to-many machine translation, achieving sublinear latency scaling with the number of target languages and supporting zero-shot transfer to unseen source languages.

WangNet is a lightweight, zero-dependency neural network that determines if a number is Numberwang in 11 languages, packaged in a 1.8 MB JSON file with simple Python inference.

Google DeepMind demonstrates the use of their 3.8 Live Extended Thinking model as a programming tutor, featuring upgraded reasoning, near real-time visual understanding, automatic detection for 97 languages, and background tool calling.

This paper systematically evaluates token merging for multilingual speech recognition on the Whisper model family, demonstrating improved computational efficiency with minimal accuracy loss across low-resource languages and fine-tuned models.

The article asks if there are better small AI models than Qwen3.5 4B for building a fast local assistant, focusing on improving capabilities like conversation, reasoning, multilingual support, and tool calling while maintaining speed.

This paper presents a three-stage training pipeline for developing compact dense retrievers, introducing PolDense for Polish and EuroDense for European languages, which achieve strong performance with significantly reduced parameters compared to larger models.

The paper introduces SWARM, a multilingual human-annotated dataset for detecting Russian propaganda in search engine results across nine languages. It benchmarks various models, finding that content-level analysis with LLMs outperforms traditional methods for propaganda detection.