This thesis introduces the Adaptive Reasoning Curriculum (ARC), a standardized framework that uses Large Language Models (LLM) for Automated Curriculum Learning (ACL). Designed to improve generalization in complex Open-Ended Learning (OEL) environments like Neural MMO 2, ARC facilitates efficient skill acquisition rather than environment memorization. Experiments demonstrate that ARC significantly outperforms expert-curricula and existing ACL methods across average return and sample efficiency. By providing open-source adapters and auxiliary tools, ARC establishes a reproducible, accessible standard for integrating LLMs into multi-agent RL, fostering broader innovation in open-ended research.

Date: 14.01.2026 / 13:00 Place: A-212

This study proposes a framework for investigating the alignment between institutional reviewer guidelines and peer review practice. A dataset was constructed from the ICLR 2024 venue on OpenReview. Large Language Model (LLM) pipelines were utilized to extract reviewer inquiries, perform topic modeling, and apply soft classification. These outputs were compared against the ICLR 2024 reviewer guideline to quantify the guideline-practice gap. Generated topics are also mapped to the guidelines from the top ten AI venues. The proposed framework provides a data-driven approach to identify implicit evaluation norms for venue organizers and gives authors and reviewers a data-driven roadmap of implicit evaluation norms.

Date: 14.01.2026 / 14:30 Place: A-212

This thesis proposes a data-driven framework to predict the surface roughness (Ra) of additively manufactured parts before printing. The most influential printing parameters were identified, and an experimental design was developed to generate a new, comprehensive dataset that contributes to the literature. Machine-learning and deep-learning models were implemented, with a multilayer perceptron (MLP) as the primary model to capture nonlinear effects. Data augmentation via a Conditional Generative Adversarial Network (CGAN) was investigated, and its impact on generalization was quantified. Finally, a custom web-based graphical user interface (GUI) was implemented for 3D model upload, interactive orientation and parameter adjustment, and real-time Ra visualization.

Date: 06.01.2026 / 13:00 Place: A-212

This thesis addresses the "tuning problem" in haptics by developing a human-in-the-loop Bayesian framework to estimate precise mechanical parameters from subjective preferences. The approach utilizes Gaussian Processes that learn from pairwise comparisons of different options to model latent utility functions. An Active Learning framework is adopted to intelligently select these comparison pairs. Experimental results across stiffness K, damping B, and distance D demonstrate that the model successfully converges to reference parameters. Furthermore, the study validates how coactive feedback accelerates convergence and helps the algorithm escape local optima.

Date: 22.01.2026 / 09:00 Place: A-212

This thesis develops a modular Value-at-Risk (VaR) framework tailored for the Borsa Istanbul (BIST) equity market. To overcome the failure of normality assumptions in emerging markets, we integrate volatility filtering with semi-parametric tail modeling. We introduce a dynamic scaling mechanism (kdyn​) that adjusts forecasts based on recent violations. Empirical analysis of 28 liquid stocks (2005–2025) shows that our adaptive Filtered Historical Simulation model achieves an 82.1% success rate in Conditional Coverage tests. These findings validate that decoupling volatility dynamics from tail shape significantly improves risk estimation stability during market stress.

Date: 12.01.2026 / 13:30 Place: A-212

Predicting surgical case durations accurately is essential for operating room scheduling and resource management of healthcare facilities. While everything relies on technology recently, enabling a system that predicts surgical durations using machine learning enhances accuracy, decreases workload of the healthcare personnel and increases patient satisfaction eventually. Using RFID-derived data to develop this machine learning model is more precise because the data source is not manually recorded. By dividing the available surgical data into three as short, medium and long, predictions are made by giving an estimation interval and the results are coherent with the literature according to the error rates.

Date: 13.01.2026 / 10:00 Place: A-212

This thesis proposes a unified framework for few-shot vision-language adaptation. Addressing the limitations of holistic image matching in models like CLIP, we introduce a dual-branch architecture combining global prompts with a locality-aware pathway. This local branch utilizes Value-Value (V-V) attention and Optimal Transport (OT) to enforce balanced, discriminative alignments between fine-grained image patches and class-specific prompts. Extensive evaluation on 11 benchmarks demonstrates state-of-the-art average accuracy. Furthermore, the framework exhibits superior Out-of-Distribution (OOD) robustness, offering a configurable trade-off between task specialization and generalized robustness.

Date: 05.01.2026 / 13:00 Place: A-212

This thesis study investigates the adoption of Generative Artificial Intelligence (GenAI) technologies among small and medium-sized IT enterprises in Türkiye. Using the Delphi method and Interpretive Structural Modeling (ISM), the research identifies key barriers and examines their interrelationships to reveal hierarchical influences. The study explores technological, organizational, and environmental determinants shaping GenAI adoption. Findings provide an empirically grounded framework that supports strategic decision-making for SMEs aiming to integrate generative AI effectively. The proposed model offers both theoretical insight and practical guidance for advancing AI-driven digital transformation.

Date: 22.12.2025 / 10:00 Place: B-116

Somatic mutations drive the transformation of normal cells into cancer. However,distinguishing driver mutations, which confer a selective growth advantage, from the vast background of neutral passenger mutations remains a critical challenge. To address this, we introduce a novel graph-based framework that constructs mutation-centric networks by leveraging longrange genomic interaction data. Our method models genomic intervals as nodes and their long-range interactions or overlaps as edges. Starting from a ’seed’ mutation, the graph expands iteratively, finding overlaps and interacting intervals to capture both local and distal genomic context. This architecture allows us to quantify a mutation’s topological influence, identify complex structural patterns (such as graph cycles), and assess proximity to known driver genes across variable ranges. Furthermore, this approach naturally generates embeddings for individual mutations, enabling the clustering of samples based on mutation profile similarity. Ultimately, by providing a comprehensive, interaction-aware view of the genomic landscape, our framework facilitates more accurate driver identification and improved patient stratification for personalized treatment.

Date: 13.01.2026 / 11:00 Place: A-212

This study presents a data-driven approach for modeling strong motion data and soil characteristics using generative AI. A Conditional Convolutional Variational Autoencoder was trained on amplitude and phase spectrograms of earthquake waveforms to generate realistic strong-motion data. The model, fine-tuned with limited site-specific data, effectively captures physical patterns and soil-dependent features without theoretical assumptions or heavy computation. Validation through fundamental site frequency analysis achieved an alignment score of 0.84, demonstrating that generated waveforms accurately reproduce site-specific frequency characteristics and distinguish between different locations.

Date: 03.11.2025 / 10:00 Place: A-212