Tag
The paper introduces TRIDENT, a round-trip translation-based framework with three classification variants, and SHAKTI, a large-scale benchmark dataset for detecting AI-generated Hindi text, evaluated across multiple genres, LLMs, adversarial settings, and seven languages.
研究了大语言模型在豪萨语等低资源语言中的安全对齐问题，提出HausaSafeLM安全偏好数据集，并发现仅靠拒绝率指标无法可靠衡量低资源语言中的模型安全性。
该论文提出一个结合人工整理词汇表与少量人类范文的框架，利用启用推理的LLM为极低资源乃至未见语言生成流畅文本，为合成数据自举提供了一条实用路径。
本文对51个近期多语言基准（涵盖242个数据集、219种语言）进行了数据驱动的审计，从覆盖度、代表性和严谨性三方面揭示了多语言评估的结构性缺陷，如英语中心化翻译、语言覆盖稀薄和基准污染问题，并提出了构建原生、文化相关且抗污染的评估框架的建议。
Researchers present the first open Speech Language Model for Basque, based on the LLaMA-Omni architecture with a Latxa backbone, using translated and TTS-synthesized data to overcome low-resource scarcity for end-to-end spoken interaction.
The paper shows that letters-only pre-tokenizer regexes in Llama-3 and Qwen3 impose an irreducible ~2.2x token premium on Nepali, and proposes an English-preserving regex repair that, with 32k new merges, reduces the premium to 1.01x, testing it via continued pretraining experiments.
This paper introduces an embedding-based subset selection method using Whisper's pretrained encoder and a greedy k-center algorithm to pick diverse training subsets for ASR fine-tuning. Evaluated on Hebrew and Danish, it shows k-center selection beats random sampling, especially at small subset sizes, due to greater lexical diversity.
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.
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。
一项覆盖14种语言的大规模研究，探讨跨语言仇恨言论检测的迁移效果，发现语言相似性是最强预测因素，而文化线索与交际规范可部分解释剩余差异。
A report examining the readiness of African languages for inclusion in AI safety research, classifying them into four readiness tiers based on data availability, national AI policy support, benchmark performance, and speaker population. It aims to guide investment and benchmark inclusion for underrepresented languages.
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.
该论文提出一个多粒度评估框架，结合字符错误率、音节错误率和分词错误率，评估藏语语音识别在多方言与不同适配策略下的表现，发现低字符错误率并不总带来更优的高层级指标，且 LoRA 适配在测试范围内仍逊于全量微调。
We propose CulturalRiddles, a benchmark of 62 cultures (45 countries, 51 languages) of human-created, culturally grounded riddles requiring abductive and compositional reasoning. Even frontier LLMs reach at most 22% accuracy on some communities, highlighting gaps in cultural understanding.
OpenAI announced model advances that significantly improve coding in low-resource programming languages like COBOL, Fortran, and Visual Basic, introducing MultiLangBench and a COBOL migration evaluation where GPT-6 Astra and GPT-6.1 Sol show major accuracy and cost gains over GPT-5.6 Sol.
OpenAI 发布新模型能力，显著提升在 COBOL、Fortran、Visual Basic 等低资源编程语言上的构建与迁移表现：GPT-6 Astra 在 MultiLangBench 20 种语言上平均通过率达 87.4%，GPT-6.1 Sol 的 COBOL 迁移准确率最高达 97.9%，且成本更低。
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.
论文提出 HeRE，一个通过 LLM 引导的知识图谱对齐去噪构建的希伯来语关系抽取基准与 150 万条银标数据集，并提出可推广至其他低资源语言的可复现流水线。研究发现多语言模型在少样本场景优于特定语言模型，而 Gemma-4-31B-it (F1=0.915) 已接近 GPT-5.4 (0.921) 的表现。
Pearl-Chat 论文介绍了一个专为低资源班图语 Luganda 设计的 8.5M 参数因果解码器 Transformer，使用 JAX/Flax NNX 从零构建，并训练了自定义字节级 BPE 分词器，实现了相比 LLaMA-3/Gemma 约两倍的压缩改进。
这篇论文通过在Tagalog上的受控实验（MorphAustro-Tok与BPE及打乱形态素的对照组比较），发现形态感知分词确实能提升机器翻译效果，但形态正确性本身并不贡献收益——增益仅来自分词粒度与一致性，而非语言学上正确的切分。