Skip to content
YTU.COSMOS

Projects

Visit the group’s project websites from here.

  • Chain-Flow

    A flow-matching hidden-state drafter for lossless speculative decoding in vLLM.

    Chain-Flow is a flow-matching hidden-state drafter for lossless speculative decoding in vLLM. Flow-Drafter runs a conditional flow over the target model's final hidden states and drafts multiple tokens in a single forward pass, then verifies with vLLM's own verifier so the output distribution matches the target exactly. With fused CUDA kernels and English/Turkish Flow-Drafter checkpoints for Qwen3.5 (4B, 9B, 27B), we report speedups up to about 2.51× in English and 2.48× in Turkish on the 27B model.

    Visit site

  • ModernBERT-TR

    A 150M-parameter Turkish encoder family with 8,192-token context window and downstream task models.

    We are releasing ModernBERT-TR. It has 150M parameters, was trained with masked-language-model pretraining, and supports 8,192-token inputs. In addition to the base model, we are releasing a family of downstream task models for aspect-polarity classification, aspect-term extraction, personally identifiable information recognition, MASSIVE intent classification and slot filling, dense retrieval, cross-encoder reranking, and guardrail. For the guardrail task we also release Guardrail-TR, a ~405K-row Turkish prompt-safety dataset.

    Visit site

  • Labyrinth

    Following Ariadne's Thread — Inducing Spatial Reasoning in LLMs Through Maze Navigation

    Ariadne is a scratch-trained language model that solves symbolic mazes through token-by-token reasoning. This interface streams every generated token and synchronizes each inference step with a visual state transition. It exposes live behavior, long-horizon spatial reasoning, backtracking strategy, and phase-level training progression in one research dashboard.

    Visit site