Sadegh Keshtkar
Biography
Since December 15, 2022, Mr. Sadegh Keshtkar has been an AI Developer in the “Computing” working group (AG C). He completed his second Master's in Computer Science at Tu Berlin and is now pursuing a Ph.D. in Computer Science at GWDG, under the supervision of Prof. Julian Kunkel. At Tu Berlin, his Master's thesis focused on studying how an automated red teaming agent learns using deep reinforcement learning. For his Ph.D., he's diving into federated learning's characteristics, exploring ways to improve this area of study with Prof. Kunkel's guidance.
ORCID: 0000-0002-4683-0136
Research Interests
- Federated learning
- Reinforcement learning
- Application of machine learning methods
Teaching
Open Thesis Topics
Federated Learning from Private LLM-Agent Trajectories for Adaptive Agent ControlApply
This thesis investigates how LLM-based agents can improve their task-solving behavior from locally generated interaction trajectories without sharing private data. During task execution, an agent produces trajectories consisting of actions, observations, tool calls, failed attempts, corrections, and final outcomes. The project studies privacy-preserving methods for using these trajectories to improve agent control while keeping the base LLM frozen or only lightly adapted. Three improvement strategies may be considered: training a lightweight next-action ranker or critic, storing and retrieving successful trajectories as contextual memory, and optionally fine-tuning small adapters such as LoRA on trajectory-derived examples. Federated learning is used to aggregate improvements across simulated private environments while raw trajectories remain local. The thesis will evaluate whether trajectory-based federated improvement increases task success, reduces unnecessary steps, and lowers execution cost compared to local-only and non-adaptive agent baselines.
FedToolAgent: Privacy-Preserving Federated Learning of Tool-Use Policies for LLM AgentsApply
This thesis investigates how LLM-based agents can learn better tool-use behavior across distributed private environments without sharing sensitive task data or execution logs. Modern LLM agents interact with external tools such as search engines, databases, code execution environments, document processors, validators, and domain-specific APIs. The project focuses on learning a lightweight tool-use policy that decides which tool should be called next, when it should be used, and how tool-call outcomes should guide subsequent actions. Each client trains locally from its own tool-use logs, including successful and failed tool calls, execution cost, latency, and task outcomes. Federated learning is then used to aggregate policy updates while keeping raw logs and private data local. The thesis evaluates whether federated tool-use learning improves task success, reduces unnecessary tool calls, lowers execution cost, and generalizes across heterogeneous client environments.
FedCollab: Federated Optimization of Collaboration Protocols in LLM-Based Multi-Agent SystemsApply
This thesis investigates how collaboration protocols in LLM-based multi-agent systems can be optimized across distributed private environments without sharing sensitive task data or interaction logs. In multi-agent systems, several specialized agents may cooperate through roles such as planner, executor, reviewer, critic, retriever, or validator. The quality of the final result depends not only on the individual agents, but also on the collaboration protocol: how tasks are decomposed, which agent acts first, how intermediate results are exchanged, how disagreements are resolved, and when the system should stop. This project studies lightweight learning methods for adapting such collaboration protocols from local multi-agent interaction logs. Federated learning is used to aggregate protocol improvements across simulated private clients while keeping raw conversations, task inputs, and agent traces local. The thesis will evaluate whether federated protocol optimization improves task success, reduces redundant communication, lowers execution cost, and increases robustness compared to fixed multi-agent collaboration patterns.
FedAgenticRAG: Privacy-Preserving Multi-Agent Reasoning over Distributed Knowledge BasesApply
This thesis investigates a privacy-preserving agentic Retrieval-Augmented Generation framework for reasoning over distributed knowledge bases without centralizing sensitive documents. In many real-world settings, relevant knowledge is distributed across different organizations, databases, departments, or user-owned collections, where direct data sharing is not possible. The proposed system uses multiple specialized LLM-based agents, such as local retrievers, summarizers, validators, and reasoning agents, to process knowledge locally and exchange only controlled intermediate representations, evidence summaries, or model updates. The work explores how federated or decentralized coordination mechanisms can support multi-step reasoning across private knowledge sources while preserving data ownership and access constraints. The thesis will evaluate the system in terms of answer quality, evidence grounding, privacy preservation, communication cost, and robustness compared to centralized and single-agent RAG baselines.
FedGuardAgent: Federated Safety Policy Learning for Autonomous LLM AgentsApply
This thesis investigates how autonomous LLM-based agents can learn and improve safety policies across distributed private environments without sharing sensitive interaction logs or task data. As LLM agents increasingly interact with tools, files, APIs, code execution environments, and external systems, they require safeguards that decide when an action is allowed, risky, should be modified, or must be blocked. The project focuses on training lightweight safety components, such as risk classifiers, policy checkers, action filters, or guard agents, from locally observed agent behavior and safety outcomes. Federated learning is used to aggregate safety-policy improvements across multiple clients while raw conversations, tool calls, documents, and execution traces remain local. The thesis will evaluate whether federated safety learning improves unsafe-action detection, reduces policy violations, preserves useful task completion, and generalizes across heterogeneous agent environments.
FedMARL-Agent: Federated Multi-Agent Reinforcement Learning for Privacy-Preserving LLM Agent CoordinationApply
This thesis investigates how multiple LLM-based agents can learn better coordination strategies across distributed private environments using federated multi-agent reinforcement learning. In complex agentic systems, several agents may interact through roles such as planner, executor, retriever, tool user, critic, validator, or summarizer. Their overall performance depends on coordination decisions, including task allocation, turn-taking, communication, conflict resolution, and stopping behavior. The project studies how local multi-agent interaction logs and reward signals can be used to train lightweight coordination policies while keeping raw conversations, task data, tool outputs, and execution traces local. Federated learning is used to aggregate policy updates across clients, enabling privacy-preserving improvement of multi-agent coordination. The thesis will evaluate whether federated reinforcement learning improves task success, reduces redundant communication, lowers execution cost, and generalizes across heterogeneous agent environments compared to fixed coordination protocols and local-only learning.
Theses
- Learning to Attack: Automated Red Teaming Using Deep Reinforcement Learning, Sadegh Keshtkar (Master's Thesis), Advisors: Sahin Albayrak, 2022-05, BibTeX
Publications
2024
- State-of-the-art artificial intelligence techniques in healthcare publications, and their correlation with disease and data: A data driven analysis (Sadegh Keshtkar, Dagmar Krefting, Anne-Christin Hauschild, Zully Maritza Ritter, Narges Lux, Aasish Kumar Sharma, Pavan Kumar Siligam, Julian Kunkel), In Journal of Artificial Intelligence and Robotics (1), 2024-11-26 BibTeX URL DOI
All publications as BibTex