- Diversity-Guided Dynamic Ensemble Selection for Multi-Label Data Streams
Mert Barkın Er
Master Student
(Supervisor:Prof.Dr.Fazlı Can) Computer Engineering Department
Bilkent University
Abstract: In multi-label data stream classification, concept drift can change both feature distributions and label dependencies over time. Existing ensemble methods either rely on fixed accuracy–diversity trade-offs or use homogeneous base learner pools. We propose ML-DynED, an ensemble method that maintains a heterogeneous pool of Multi-Label Hoeffding Trees, Multi-Label k-Nearest Neighbors, and Binary Relevance classifiers, and selects an active sub-ensemble using a multi-label adaptation of Maximal Marginal Relevance. In experiments on 30 real-world and synthetic datasets, ML-DynED achieves the best micro-F1 on several benchmarks, with an average rank of 13.29 across 32 methods. Although it does not consistently outperform the strongest instance-based baselines, the results suggest that diversity-aware selection is a competitive alternative to homogeneous ensemble design for multi-label data streams.
DATE: March 23, Monday @ 15:30 Place: EA 502
2. Contact-Rich Adaptive LLM-based Control for Robotic Manipulation
Mert Kaan Er
Master Student
(Supervisor:Asst.Prof.Özgür S.Öğüz) (Co-advisor:Assoc.Prof.Hamdi Dibeklioğlu) Computer Engineering Department
Bilkent University
Abstract: While Large Language Models (LLMs) and Vision-Language Models (VLMs) demonstrate capabilities in reasoning and semantic understanding, applying them directly to contact-rich manipulation remains a challenge due to their lack of physical grounding and adaptive control. We propose CoRAL (Contact-Rich Adaptive LLM-based control), a modular framework enabling zero-shot planning by decoupling high-level reasoning from low-level control. Unlike black-box policies, CoRAL utilizes LLMs not as direct controllers, but as cost designers that synthesize context-aware objective functions for a sampling-based motion planner (MPPI). To address visual ambiguity in physical parameters, a VLM first provides semantic priors for environmental dynamics (e.g., mass and friction). The LLM then refines these estimates using interaction history and iteratively updates the cost function structure to correct strategic errors. A retrieval-based memory unit allows the system to reuse successful strategies across recurrent tasks. This architecture ensures real-time control stability, bridging the gap between slow LLM inference and dynamic contact requirements. We validate CoRAL in simulation and on real-world hardware across challenging tasks. Experiments demonstrate that CoRAL outperforms state-of-the-art VLA and foundation-model-based planner baselines by boosting success rates over 50% on average in unseen contact-rich scenarios, effectively handling sim-to-real gaps through its adaptive physical understanding.
DATE: March 23, Monday @ 15:50 Place: EA 502
3. INSPECTBUGS: Detecting Invalid Bug Reports and Resolving with No-Code Fixes
Mahmut Furkan Gön
Master Student
(Supervisor:Assoc.Prof.Eray Tüzün) Computer Engineering Department
Bilkent University
Abstract: Many software issues are reported to software maintainers in the form of bug reports. However, there are many invalid bug reports whose solutions do not require modification in the source code. The existence of invalid bug reports causes unnecessary human effort and time spent in determining their nature. Also, customer support staff spend a considerable amount of time explaining why the reported bug is invalid. In this study, we investigate the automated subclassification of invalid bugs by utilizing different machine learning (ML), deep learning (DL), and large language model (LLM) techniques. We also study how subclassification, LLMs, and retrieval-augmented generation (RAG) systems can be useful in suggesting no-code fixes to invalid bug reports. On a dataset of bug reports from the Brave repository, we evaluated various models, including ML, DL, and LLMs, to determine the subclass of invalid bug reports. We then used different LLMs in a RAG setting and standalone to generate no-code fixes for these invalid reports. We applied different evaluation mechanisms for the no-code fixes. Firstly, we randomly sampled and manually evaluated the suggested no-code fixes to see if they resolved the bug. Then, we compared the semantic similarities of the suggested no-code fixes with ground truth no-code fixes—obtained from the dataset via BERTScore. This study shows that LLM solutions outperform traditional ML and DL-based solutions, as well as stand-alone LLMs, in bug report subclassification tasks. RAG-based no-code fix suggestions successfully resolved a substantial proportion of the invalid bug reports, showing their potential for providing accurate and quick solutions to frequently reported issues. RAG-based no-code fix suggestions displayed a high similarity with real-world scenarios, showing their potential to be used in customer support services. Further analysis is needed to evaluate the effect of different fine-tuning, prompt, and context engineering techniques, as well as other LLMs and AI models in the future.
DATE: March 23, Monday @ 16:30 Place: EA 502
4. StyleFusion360: View-Consistent Head Stylization via Adaptive Style Modulation
Furkan Güzelant
Master Student
(Supervisor:Asst.Prof.Ayşegül Dündar Boral) Computer Engineering Department
Bilkent University
Abstract: 3D head stylization enables expressive reimagining of human faces for creative visual experiences in digital media. Existing 3D-aware methods often require computationally intensive optimization or per-style fine-tuning, limiting flexibility and user control. To overcome these challenges, we introduce StyleFusion360, a diffusion-based framework for multi-view consistent, identity-preserving 3D head stylization from a single style reference image, without per-style training. Our approach enhances the Style Fusion Attention mechanism with a style-conditioned key modulation mechanism that aligns content and style representations for fine-grained and controllable stylization. We further provide a user-controllable slider for adjusting stylization intensity. In addition, StyleFusion360 supports local multi-edit stylization, enabling targeted edits such as modifying hair or eyes independently. Extensive experiments on FFHQ and RenderMe360 demonstrate that StyleFusion360 produces high-quality, controllable, and visually compelling stylizations, outperforming state-of-the-art GAN- and diffusion-based methods across diverse style domains.
DATE: March 23, Monday @ 16:50 Place: EA 502