The Autonomous Agents Lab studies efficient AI systems that perceive, reason and act, alongside foundation models for scientific discovery. Our research spans tool-using language agents, multimodal perception and autonomous systems, and protein and enzyme design. These areas share a focus on learning useful representations and building models that work under practical data and computational constraints.

Efficiency and large-scale computing connect our research. We develop methods that require fewer labels, less compute and less memory, and build distributed training pipelines for large models on HPC supercomputers. This combination supports research from compact models for mobile devices to foundation models trained on large scientific and multimodal datasets.

Research Areas

Efficient language models and tool-using agents

We study language-model assistants that use tools and APIs, including compact agents for mobile devices and reinforcement learning with verifiable multi-step rewards. Related work investigates multi-agent reasoning over scientific knowledge graphs and methods for retrieving and using external knowledge.

Multimodal perception and autonomous systems

We develop methods for combining information from cameras, lidar, radar and thermal sensors. Our research includes multimodal language models, dense correspondence, open-vocabulary 3D object detection and all-weather autonomous driving, with autonomous racing on the F1TENTH platform providing an experimental setting.

Foundation models for protein and enzyme design

We study protein language models trained on metagenomic sequences and generative methods for industrial enzyme design. Projects include metagenomic functional annotation and active-learning-guided directed evolution, connecting computational modelling with experimental enzyme research.

Collaborative applications

With partners, we apply these methods to aerospace defect detection, manufacturing cost prediction, lung cancer survival prediction, historical-source question answering, conflict harm documentation, recommendation, survey pretesting and financial forecasting.

People

Principal Investigator

Asst. Prof. Taha Koçyiğit received his PhD from the University of Edinburgh, where his thesis focused on compute- and data-efficient training of computer vision models, and his MSc from the Technical University of Munich. (Google Scholar)

PhD Students

• Nesibe Şebnem Paluluoğlu: discovery, design and active-learning-guided directed evolution of ketoreductases for the stereoselective reduction of tropinone (co-supervised with Asst. Prof. İbrahim Çağrı Kurt)
• Ahmet Yasin Aytar: a multi-agent framework for forecasting emerging research directions over temporal scientific knowledge graphs (co-supervised with Assoc. Prof. Şener Özönder)
• Gökhan Yıldırım: reinforcement learning for tool-using language agents with verifiable multi-step rewards (co-supervised with Prof. Şuayb Arslan)

MS Students

• Fatma Nur Dumlupınar Keşir: multimodal recommendation with knowledge-graph retrieval and LLM-based reranking
• Ömer Faruk Kolçak: building a metagenomic functional atlas with language models (co-supervised with Asst. Prof. İbrahim Çağrı Kurt)
• Manolya Çetinkaya: multimodal large language models (co-supervised with Prof. Şuayb Arslan)
• Mahmut Enes Akten: pretesting surveys with synthetic respondents, an empirical evaluation of large language models
• Doğanay Tuhan: predicting stock prices using macroeconomic data and sentiment analysis (co-supervised with Asst. Prof. Mustafa Metin Başbay)

Alumni

• Ömer Faruk Deniz (MSc, 2026): Open-Vocabulary 3D Object Detection with Promptable Segmentation

Ongoing Projects

• On-Device LLM Assistant for Real-Time and Dynamic API Execution ("Cihaz Üzerinde Gerçek Zamanlı ve Dinamik API Yürütümü için LLM Asistanı"). Principal Investigator. Boğaziçi University BAP, YÖK ADP, 2025–2027.
• A Frontier-Scale Protein Language Foundation Model (EHPC-AIF-2025LS17-013). Co-Principal Investigator. EuroHPC, Horizon Europe, 2026–2027.
• Cross-API Agents for Reliable End-to-End Workflow Automation (EHPC-AIF-2025LS17-020). Co-Principal Investigator. EuroHPC, Horizon Europe, 2026-2027
• The Dissemination of Knowledge in the Classical Islamic World: Rihlas (123K484). Researcher. TÜBİTAK 1001, 2023–2026.

Completed Projects

• Multi-Modal Large Language Model for Cross-Modal Reasoning and Generation through Scalable Dataset Annotation and Contrastive Alignment (EHPC-DEV-2025D06-088). Principal Investigator. EuroHPC, Horizon Europe, 2025–2026.
• Robust Dense Correspondence via Multi-Modal Fusion and Transformers ("Çok Modlu Birleştirme ve Transformer'lar ile Sağlam Yoğun Eşleştirme"). Principal Investigator. Boğaziçi University BAP, 2025–2026.
• On-Device AI Agents: Revolutionizing Real-Time API Interactions with Privacy-Preserving Intelligence (EHPC-AI-2024A04-084). Co-Principal Investigator. FFplus and EuroHPC, Horizon Europe, 2025.
• Revolutionizing Industrial Enzymes through Generative AI ("Üretken Yapay Zekâ ile Endüstriyel Enzim Devrimi", EHPC-AI-2024A04-095). Co-Principal Investigator. EuroHPC, Horizon Europe, 2025.
• Scalable Multi-Modal AI for Cross-Modal Reasoning (EHPC-BEN-2024B12-040). Co-Principal Investigator. EuroHPC, Horizon Europe, 2025.
• Multi-Modal End-to-End All-Weather Autonomous Driving ("Çoklu-Modalite Uçtan-Uca Tüm Hava Koşullarında Otonom Sürüş"). Co-Principal Investigator. TÜBİTAK BİLGEM, 2024–2025.
• AI-Based Cost Estimation Software for Automotive Suspension and Steering Parts ("Otomotiv Süspansiyon ve Direksiyon Parçaları için Yapay Zekâ Tabanlı Maliyet Tahmin Yazılımı"). Researcher. TÜBİTAK TEYDEB 1711, 2024–2025.

Publications

1. Ö. F. Deniz, M. T. Koçyiğit. Open-Vocabulary 3D Object Detection with Promptable Segmentation. arXiv:2609.19358, 2026. [paper]
2. M. Gökpınar, Y. Almalioglu, M. T. Koçyiğit, T. Kahveci, D. Demir, K. Başak, M. Turan. Vision Transformer Supported Kolmogorov-Arnold Networks for Survival Prediction in Lung Cancer. ACM Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM BCB), 2025. [paper]
3. A. Metiner, Y. Nikishkov, A. Makeev, M. T. Koçyiğit. Deep Learning-Enhanced X-Ray Computed Tomography for Defect Detection in Composite Structures. Journal of Nondestructive Evaluation, 44(4):127, 2025. [paper]
4. A. B. Arıkan, Ş. Özönder, M. T. Koçyiğit, H. O. Altun, H. K. Küçükkartal, M. Arslanoğlu, F. Çağırankaya, B. Ayvaz. Machine Learning-Based Manufacturing Cost Prediction from 2D Engineering Drawings via Geometric Features. arXiv:2508.12440, 2025. [paper]
5. N. Ş. Paluluoğlu, D. Z. Gürer, M. F. Akıncı, M. T. Koçyiğit. GazaVHR: AI-Driven Legally Grounded Conflict Harm Documentation. 4th Muslims in ML Workshop at ICML, 2025. [paper]
6. A. Metiner, Y. Nikishkov, M. T. Koçyiğit, A. Makeev. Application of Deep Learning for Defect Detection by X-Ray Computed Tomography in Large Aerospace Structures. AIAA SciTech Forum, 2025. [paper]

Join Us

We welcome MSc and PhD students interested in efficient AI, high-performance computing, language-model agents, autonomous systems or AI for protein design. Applications are made through the institute's graduate admissions. Before applying, feel free to email taha.kocyigit@bogazici.edu.tr with your CV and a few sentences on the topic you would like to work on.