Journal paper under revision
ModernBERT-TR
A long-context (8,192 tokens) Turkish encoder with a family of specialized downstream models.
Research
Three threads: building the models, teaching them to reason and specialize, and making them run where they're needed.
Building the models
From large clean Turkish corpora to long-context encoders and multimodal models.
Teaching them to reason and specialize
Training small LLMs with logic games and maze navigation; domain adaptation.
Making them run
New training and inference methods, such as flow-matching speculative decoding, and deployment to edge devices.
Selected research
Best Paper Award, UBMK 2025
Reasoning with logic games
Training small LLMs to reason better through games and structured logic-based tasks. Best Paper Award, UBMK 2025.
Publications
0 papers · 0 first-author · 0 co-first (equal contribution)
* Equal contribution
Several of these works were developed with YTU COSMOS AI Research Lab. YTU COSMOS lab →