We introduce a permutation-invariant model that reconstructs a receiver-centred vector floor plan from any 4–16 irregularly positioned, position-tagged single-channel room impulse responses (RIRs). A single checkpoint improves from 86.3% mean intersection over union with four measurements to 96.2% with sixteen measurements, allowing acquisition cost, spatial coverage, and reconstruction accuracy to be traded off at deployment without retraining. Random measurement subset training simultaneously enables variable-cardinality inference and serves as structured measurement dropout regularization. The model returns an ordered vector contour and is evaluated on a 240,000-room general-polygon benchmark spanning 4-8 corners, up to five reflex vertices, and certified first-order-invisible walls.

Date: 15.09.2026 / 10:00 Place: A-212

This thesis investigates whether spatially controlled auditory cues can mitigate spatial disorientation (SD), a vestibular illusion common in flight. Twenty participants underwent Barány-chair yaw rotations while visually blocked, with sound delivered via an eight-speaker ring under congruent, incongruent, static, and silent conditions, using three sound types. Multimodal measures (EOG-derived nystagmus, EEG, fNIRS, eye tracking, ECG, joystick) revealed an asymmetric effect: congruent, rotation-matched audio amplified the vestibular response, especially pseudo-speech, while incongruent and static sounds failed to reduce disorientation below the silent baseline. Rather than countering SD, moving sound intensified it, clarifying which configurations do not work and informing future countermeasure design for pilot training.

Date: 04.09.2026 / 10:00 Place: A-212

This thesis investigates both the structural and mechanistic interpretability of OpenVLA. Structural interpretability is examined through conventional linear probing, which maps the hierarchy of representations, whereas mechanistic interpretability is investigated using a log-probability-based probing framework that analyzes the causal use of information through the early-exit mechanism. Both methods are applied layer by layer using task-specific controlled perturbations. By comparing correlation-based and causality-based analyses, the proposed framework provides a comprehensive understanding of how autoregressive Vision-Language-Action models represent and utilize linguistic and action-related information.

Date: 02.09.2026 / 15:00 Place: B-116

The anatolian leopard is a scarcely studied, possibly extinct population of the leopard subspecies tulliana. In this thesis, we sequence potential big cat genomes from a wide temporal and spatial distribution across Anatolia and obtain three workable quality Panthera pardus genomes. We construct mitochondrial and whole-genome phylogenies of the Anatolian leopard. The study represents the first ever whole-genome study into the Anatolian leopards.

Date: 03.09.2026 / 16:00 Place: A-212

This thesis compares generative and encoder-based language models for multilingual phishing email detection on local, resource-constrained environments. Utilizing the synthetic Turkish OLTA-TR dataset built on Cialdini's persuasion principles, we train models to identify psychological manipulation rather than static keywords. Experimental results on MeAJOR and OLTA-TR demonstrate that while generative models exhibit latency issues and formatting instability due to a helpful AI bias, fine-tuned classification models achieve superior accuracy, highlighted by mmBERT's 99.49% Macro F1-Score. A dual-layer interpretability framework is also introduced to explain model decisions, concluding that classification models provide more secure real-time defense.

Date: 02.09.2026 / 10:00 Place: A-212

This thesis introduces a complexity measure for four UML diagram types, so that model performance can be analysed by difficulty level. It also builds an evaluation framework that scores a generated diagram against its reference automatically. Six locally deployable models and a 70B frontier model were run on class, sequence, state, and activity diagrams. The best local model was then improved through complexity-aware few-shot in-context learning. Each request was paired with an example matched to the complexity level of the target diagram. This lifted the model above the 70B model on three of the four types.

Date: 26.08.2026 / 10:00 Place: A-108

This thesis examines the operational performance of incident investigations in home health care, using routinely collected records from a US home healthcare provider (July 2023–December 2025; 493 mandatory investigations). Prolonged investigations were defined through an empirical 80th-percentile threshold (≥15 days), classifying 20.1% of cases. The prolonged rate remained statistically stable over the 29-month period. Nested logistic regression showed that incident category did not explain prolongation, whereas the organizational unit significantly improved model fit; county-level rates ranged from 0% to 58.3%. The study offers a transferable, data-driven approach for monitoring investigation performance without additional data collection.

Date: 03.09.2026 / 10:30 Place: A-212

Technical documentation for enterprise electronic products spans several formats such as datasheets and brochures, and locating required information is challenging. This thesis proposes a graph-guided RAG architecture for question answering over electronic product documentation, modelling product relationships and attributes in a Neo4j knowledge graph and filtering queries through this graph before vector search. A rule-based intent router directs each query to the appropriate processing flow, and a tiered chunk-gating mechanism enforces this scope during retrieval. Compared against a hybrid RAG baseline, the approach reduces cross-product contamination by 78%, with statistically significant improvements in retrieval accuracy and answer quality.

Date: 31.08.2026 / 13:00 Place: A-212

This thesis presents CBC-Distilled, an efficient multimodal CNN–Transformer model for remote sensing semantic segmentation under missing-modality conditions. The model combines RGB, near-infrared (NIR), and shortwave infrared (SWIR) information and uses feature distillation, cross-band correlation, and random modality masking to improve robustness when one or more modalities are unavailable. The teacher and auxiliary branches are used only during training and removed during inference to reduce computational cost. Experiments show that the proposed model maintains competitive segmentation performance while requiring fewer parameters and lower computational cost than the baseline.

Date: 03.09.2026 / 13:30 Place: A-212

Continuous production environments (refineries, power grids) generate vast multivariate sensor data, yet remain stuck in fragmented “Narrow AI” while foundation models transformed language and vision. This thesis introduces Prometheus, a compact 1.6M-parameter bidirectional Transformer encoder that reads a plant’s sensors as a language and learns its coupled physics through self-supervised “Four-Teacher” geometric masking. On a crude distillation unit, Prometheus wins all 38 channels on every metric, surpasses zero-shot Time-Series Foundation Models up to ~300× larger, reconstructs entirely missing sensors, and beats a deployed industrial soft sensor, evidence that domain specialization, not generalized scale, is the path toward Industrial Foundation Models.

Date: 25.06.2026 / 14:30 Place: A-212