Graduate School of Informatics

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

Large language models often answer confidently even when they lack relevant knowledge. This thesis asks whether they can learn a crucial form of epistemic humility: recognizing and admitting the limits of what they know. It introduces hypothetical terms, plausible concepts screened across multiple sources and likely absent from training data, to measure how readily models invent unsupported explanations. The resulting benchmarks reveal confabulation across model families, sizes, and reasoning systems. Targeted fine-tuning reduces this tendency and improves factuality across three architectures. Finally, the thesis characterizes how learned uncertainty is represented inside models and when it shapes their responses.

Date: 31.08.2026 / 14:00 Place: A-212

This thesis proposes a cyber resilience orchestration framework that extends security orchestration beyond conventional detection and response. The framework integrates Large Language Model (LLM) agents, expert-validated playbooks, automation readiness controls, and API recommendation mechanisms to support automated and human-controlled resilience operations. Experimental evaluations demonstrate the feasibility of the proposed approach under controlled conditions.

Date: 03.09.2026 / 15:00 Place: A-108

This thesis develops grounded, explainable scene understanding for sensitive-content moderation. It introduces SenBen, a benchmark of 13,999 movie frames annotated with scene graphs and sensitivity tags, and SenBen-Score, a recall-oriented metric. A vocabulary-aware distillation framework transfers Gemini annotations to a compact 241-million-parameter Florence-2 model using suffix-based object identity, Vocabulary-Aware Recall loss, and a separate sensitivity-tag head. The resulting model outperforms evaluated vision-language models except Gemini and all tested commercial safety APIs on the composite benchmark, while running 7.6 times faster and using 16 times less GPU memory than the strongest competing local model.

Date: 01.09.2026 / 11:00 Place: A-212