XMuR: eXplainable MUltilingual Reasoning.

XMuR (eXplainable MUltilingual Reasoning) investigates how AI systems reason across different languages while making their decisions transparent and understandable. The project aims to develop methods for evaluating and improving multilingual reasoning, identifying linguistic biases, and producing clear explanations that users can trust.

Objectives

• Categorize reasoning types: Identify and organize the main forms of reasoning required across languages, such as logical, causal, numerical, commonsense, and relational reasoning.
• Evaluate multilingual performance: Measure how accurately and consistently LLMs perform these reasoning tasks across languages, language families, and resource levels.
• Identify performance differences: Determine how linguistic structure, training-data availability, prompting language, and cultural context influence reasoning outcomes.
• Explain LLM processing differences: Analyze the models’ internal representations and reasoning pathways to explain why LLMs process equivalent problems differently across languages.

Funding & Support

This project is financed by KUNUMI Labs, covering the period from August 2026 to December 2027.