The field of natural language processing is moving towards developing more effective and efficient multilingual large language models (LLMs). Recent research has focused on improving the performance of LLMs on low-resource languages, exploring the internal mechanisms of these models, and developing novel approaches for multilingual retrieval-augmented generation. A key direction is the investigation of how LLMs encode and retrieve knowledge across languages, with studies revealing that knowledge is encoded in a language-independent concept space but often transitions to language-specific spaces in the final layers. This has led to the development of methods that bypass computations in the final layers to enhance prediction accuracy and cross-lingual consistency. Additionally, researchers are exploring the use of dialectic reasoning and argumentations to make retrieval-augmented generation more analytical and critical. Noteworthy papers include:

• Scaling Test-time Compute for Low-resource Languages, which introduces English-Pivoted CoT Training to improve reasoning in low-resource languages.
• Improving Multilingual Retrieval-Augmented Language Models through Dialectic Reasoning Argumentations, which proposes a modular approach guided by Argumentative Explanations to systematically evaluate retrieved information.
• Lugha-Llama, which adapts LLMs to low-resource African languages by combining curated data with high-quality English educational texts.