<![CDATA[Background: Multilingual large language model evaluation increasingly recognises Indonesian as a significant benchmark language, yet Indonesia’s wider linguistic ecology requires assessment across local languages whose digital representation remains uneven. Objective: This study examines whether large language models reason consistently across Indonesian, Javanese, Sundanese, and Buginese when semantically aligned reasoning tasks are presented in each language. Method: Using a controlled multilingual evaluation design, this study compares model outputs through three analytic dimensions: final-answer consistency, explanation coherence and faithfulness, and language-linked error patterns. Results: Findings show that answer stability is not evenly preserved across language pairs, with stronger alignment in Indonesian–Javanese and Indonesian–Sundanese comparisons than in Indonesian–Buginese comparison. Explanation analysis further indicates that apparently correct answers may be accompanied by compressed, partially faithful, or weakly grounded reasoning. Implication: Error analysis reveals that inconsistency emerges through multiple pathways, including lexical-semantic drift, register mismatch, translation-related distortion, cultural inferential misreading, and language fallback. Novelty: The novelty of this study lies in shifting Indonesian multilingual LLM evaluation from isolated benchmark accuracy to cross-linguistic reasoning consistency, positioning local languages as central analytic sites for assessing reliability, equity, and epistemic accountability in multilingual artificial intelligence]]>
Copyrights © 2026
