low-resource-language

The paper presents Saidzoda-Tajik-LFM-2.6B, a Tajik adaptation of Liquid AI's LFM2.5-2.6B-Base for tool calling on modest hardware, using LoRA continued pretraining followed by supervised fine-tuning. Tajik tool-call accuracy jumps from 20.69% to 64.14%, but the authors show CPT wipes out native tool-call formats and harms English MMLU (60.8 to 52.5).

The paper introduces Latin Modern BERT (LAMB), a ModernBERT-based encoder pretrained on over 74B Latin tokens with morphologically informed masking, alongside two new Latin corpora (Acervus and Corpus Trinum), and shows it outperforms 15 existing encoders on 16 of 17 benchmark subtasks.

介绍 BenCoDiC——一个大规模孟加拉语政治新闻语境下的网络欺凌语料库，包含来自 100 个 YouTube 视频的 50,250 条文本实例，保留了字幕-评论-回复关系，并通过两级人工标注达到 Fleiss' κ=0.86 的一致性；实验表明加入对话语境能带来一致但幅度较小的分类性能提升。

AraGenre 2026 is a shared task for hierarchical, definition-guided Arabic genre classification aimed at improving annotated data in low-resource languages, with results showing strong broad-genre recognition but gaps in fine-grained classification.

This paper introduces MudawanSn, a gold-standard parallel corpus for Wolof-Arabic machine translation, and shows that fine-tuning on it yields substantial improvements in translation quality for both language directions.

NepKANUN is an AI-powered legal assistant for Nepali using RAG and a fine-tuned LLaMA 3.2 3B model to provide accurate answers to legal queries, validated by expert reviews.

This paper proposes an end-to-end sequence-to-sequence approach for contextual Tamil spelling and grammar correction, using progressively fine-tuned mT5 and mBART models on synthetic data, achieving 69.3% exact-match accuracy on a diagnostic set.

This paper studies supervised fine-tuning and reinforcement learning for reasoning in low-resource languages, revealing that accuracy benchmarks are noisy while SFT builds language-specific reasoning and RL fixes format and leakage issues.

This paper investigates fine-tuning MoE models to reason in Greek, revealing that accuracy metrics are noisy, while supervised fine-tuning and reinforcement learning improve language-specific reasoning and fix behavioral defects, with proposed evaluation instruments.

BengaliMCQ is a structure-aware RAG framework using graph neural networks to model hierarchical document structures in Bengali textbooks, enabling automatic generation and answer prediction of academic multiple-choice questions with improved performance over baseline methods.

Introduces BavGround, a benchmark for evaluating LLMs' regional cultural grounding and dialect competence in Bavarian across English, German, and Bavarian, finding that models struggle with dialectal and localized cultural knowledge.

A two-dialect finite-state morphological analyzer for the Dungan language is presented, with a multi-genre evaluation measuring inflection, ambiguity, and lexical coverage.

Mwando is a virtual educational assistant leveraging AI to preserve and teach the Comorian language shiKomori, utilizing a multi-agent architecture with vector search, knowledge graph, and web fallback.

Introduces PatiGonit22K, an expanded Bengali mathematical word problem dataset with 22,441 problems, including complex multi-operation problems, to advance mathematical reasoning research for low-resource languages.

This paper introduces KyrgyzLLM-Bench, a benchmark suite for evaluating large language models in the Kyrgyz language, comprising both natively authored and translated datasets, and provides a systematic evaluation of 26 models.

Introduces BaFCo, a benchmark dataset for Bangla form comprehension focusing on Document Layout Analysis (DLA) and Key Information Extraction (KIE). It includes 200 multi-page complex Bangladeshi government forms with fine-grained annotations across 26 entity types and evaluates multiple MLLMs, revealing limitations in understanding complex Bangla forms.

This paper introduces BanglaMemeEvidence, a multimodal dataset of 2,917 Bengali memes annotated for explanatory evidence detection, and proposes BengaliMemeEvidenceNet, a hybrid framework achieving an F1 score of 0.74.

This paper investigates using text-to-speech (TTS) to generate synthetic training data for spoken question answering in Luxembourgish, a low-resource language, and evaluates multi-source TTS configurations with a parameter-efficient SLAM-style architecture.

Presents an open-source ASR system for assessing children's reading in Bambara, including field data collection, benchmark construction, model adaptation, and classroom validation, achieving significant word error rate reduction.

This paper introduces a Bangla event detection benchmark with noisy text (ASR, orthographic corruption) and evaluates encoder-only and decoder-only LLMs, finding decoder models more robust to noise.