Research archive
Research papers
Each page includes verified metadata, a complete abstract, citation details, and links to research artifacts.
Preprint
July 16, 2026
DOI: 10.21203/rs.3.rs-10384535/v1
Ahmet Rıfat Öztürk, Yağız Ekrem Dalar, Ömer Faruk Aksoy, Nedim Mutlu
Sezer, and Feyzi Arda Salihoğlu
A version-aware, low-cost workflow for diagnosing external evaluation
failures, auditing answer-label bias, and reporting partial model repair.
Turkish NLP
Small language models
Model repair
Benchmark auditing
Read the technical report →
Anonymized manuscriptJune 5, 2026Turkish DocVQA
Anonymous authors (double-blind manuscript)
A synthetic dataset of 15,000 Turkish business-document images and 235,000 question-answer pairs, with Donut, PaliGemma-3B, and Pix2Struct comparisons.
Turkish DocVQASynthetic dataDocument understandingVision-language models
OpenReview preprint
OpenReview submission
Goal drift and AI alignment
Ethosoft research team
A research paper examining goal drift in capable AI systems and the implications of revising, preserving, and aligning objectives under changing task context.
AI alignmentGoal specificationAgentic systemsSafety
Preprint
July 14, 2026
DOI: 10.21203/rs.3.rs-10331341/v1
Ahmet Rıfat Öztürk, Yağız Ekrem Dalar, Ömer Faruk Aksoy, Nedim Mutlu
Sezer, and Feyzi Arda Salihoğlu
A controlled study of the gap between learning a sentence-level future
representation and using that signal to change language-model generation behavior.
Small language models
Future prediction
Semantic reranking
Model interpretability
Read the research summary →
Zenodo preprint
July 31, 2026
DOI: 10.5281/zenodo.21718745
Nedim Mutlu Sezer, Yağız Ekrem Dalar, Feyzi Arda Salihoğlu, Ömer Faruk Aksoy, and Ahmet Rıfat Öztürk
A long-horizon test of whether causal surprisal can identify positions where additional computation has useful marginal value.
Adaptive computationSurprisalCausal routingMamba-2
Workshop accepted · Journal under review
July 9, 2026
DOI: 10.5281/zenodo.21275744
Nedim Mutlu Sezer, Ömer Faruk Aksoy, Ahmet Rıfat Öztürk, Feyzi Arda Salihoğlu, and Yağız Ekrem Dalar
MYCELIUM tests whether safety mechanisms remain behaviorally load-bearing under removal, wrong-route, family-zero, and mandatory-path controls.
Language-model safetyDependabilityCounterfactual evaluationReproducibility
Zenodo preprint
July 31, 2026
DOI: 10.5281/zenodo.21718516
Nedim Mutlu Sezer
An EpiLang case study of how epoch endpoints select degenerate output policies and hide the difference between memorization and rule learning.
Training orderAlgorithmic generalizationMemorizationExecutable semantics