Seven open-weight Turkish LLMs evaluated on six benchmarks: knowledge, reasoning, reliability and hallucination detection. Every model sees the same items. Every score carries a 95% Wilson interval, and every comparison is backed by a paired McNemar test, computed live from per-item results.
Primary metric per benchmark, in %. Click a column header to sort. The best score in each column is bold; the small numbers are the 95% Wilson CI.
Corrected ARC-TR re-scores 20 items whose digit answer keys are 1-based. Three models did not store predictions, so a range is shown for them. See Known issue.
Scores with 95% CIs, and which pairwise differences are statistically significant (McNemar, p<0.05).
Pick two models. "Only A" counts the items A gets right and B gets wrong. The McNemar test uses only these discordant items.
@inproceedings{yilmaz2026multiaxis,
author = {Yılmaz, Muhammet Tarık and Pala, Muhammed Ali and Çakır, Yusuf},
title = {A Multi-Axis Comparative Evaluation of Turkish Large Language Models
across Knowledge, Reasoning, and Reliability},
booktitle = {2026 Innovations in Intelligent Systems and Applications Conference (ASYU)},
publisher = {IEEE}, year = {2026},
doi = {10.1109/ASYU71790.2026.11772376}
}