About Research Papers Experience Projects Reports Awards Leadership Media Contact Blog

PhD Student · EPFL IC · EDIC

Yusuf Kesmen

PhD student in Computer Science at EPFL,
building trustworthy LLMs for high-stakes medicine.

LLMs & Post-training Medical AI Trustworthy AI Agents / RAG Reasoning
Blog
Scroll
01 — About

Background


Yusuf Kesmen
Location Lausanne, CH
Languages TR · EN · AR
Program EDIC PhD

I am a PhD student in Computer Science at EPFL, admitted to the EDIC Doctoral Program with an EPFL Fellowship. Previously, I graduated as valedictorian (ranked 1st) from Bilkent University's Computer Science program with a 3.96/4.00 GPA.

My research is on building trustworthy LLMs for high-stakes medicine. The thread running through it is separating knowledge representation from inference machinery, and grounding the latter in formal frameworks from statistical decision theory. I am open to supervising semester projects and BSc/MSc theses on reasoning, post-training, and medical AI.

Beyond CS, I draw inspiration from philosophy and economics — disciplines that shape how I think about decision-making under uncertainty, the ethics of autonomous systems, and the societal impact of AI.

Education

Ph.D. in Computer Science — EPFL

School of Computer and Communication Sciences (IC) · Sep 2025 — Present

EDIC Doctoral Program · EPFL Fellowship

B.S. in Computer Science — Bilkent University

Ankara, Turkey · Sep 2021 — Jun 2025

GPA: 3.96/4.00 · Valedictorian (1st/class)

Download CV
02 — Research

Current Focus


Clinical decisions need to be both reliable and auditable — properties that today's LLMs struggle to provide on their own. The directions below are the threads I'm actively pursuing, and the ones I'd be glad to supervise as semester projects or BSc/MSc theses.

Ongoing & Proposed Project Directions

Structured Reasoning & Agentic Frameworks

Reliable AI via structured reasoning paradigms and agentic architectures with explicit, auditable inference steps.

Post-training for Reasoning & Clinical Alignment

Aligning LLMs for clinically grounded reasoning through fine-tuning, RL, and verifier-guided objectives.

Test-time Scaling & Inference-time Reasoning

Inference-time methods — search, sampling, verification — for medical decision-making under uncertainty.

Read the Research Proposal
03 — Papers

Publications


Preprint

MoBayes: A Modular Bayesian Framework for Separating Reasoning from Language in Conversational Clinical Decision Support

Y. Kesmen, F. Elhassan, J. Ma, J. Stalhandske, Y. Chang, D. Sasu, A. Kulinkina, A. Arora, L. Klein, M.-A. Hartley

arXiv:2604.20022 · May 2026
Published

From Physics to Machine Learning and Back: Part II — Learning and Observational Bias in PHM

O. Fink, I. Nejjar, V. Sharma, K. Faghih Niresi, H. Sun, H. Dong, C. Xu, A. Wei, A. Bizzi, R. Theiler, Y. Tian, L. Von Krannichfeldt, Z. Ma, S. Garmaev, Z. Zhang, M. Zhao, K. Steiner, Y. Kesmen

Contributed: Reinforcement Learning section

Reliability Engineering & System Safety, 274 (2026) 112376 · Feb 2026
Preprint

Effective Communications in the Goal-Oriented Semantic Signal Processing Framework: A Case Study in Cooperative Driving

Y. Kesmen*, Y. Küçük*, M. Kalfa, O. Arıkan

* Equal Contribution
04 — Experience

Research & Professional


Research Experience

Feb 2026 — Present

Doctoral Researcher

LiGHT Lab — EPFL · Prof. Mary-Anne Hartley

Working on LLM reasoning for high-stakes clinical domains. Developing a modular diagnostic dialogue architecture that integrates LLM-based language processing with deterministic Bayesian reasoning for auditable, reliable clinical decision-making. Also building a collaborative multi-agent framework that treats LLM reasoning as an editable sequence, refining shared reasoning chains via atomic operations and structured debate.

Sep 2025 — Jan 2026

Doctoral Researcher

IMOS Lab — EPFL · Prof. Olga Fink

Built an agentic root cause analysis framework that structures LLM reasoning through explicit hypothesis-driven diagnostic reasoning with log-odds belief tracking. Maintained belief scores over a discrete fault space, accumulating evidence across diagnostic steps for cyber-physical systems. Also contributed to a computational framework for mechanical bit readout through inverse dynamic programming.

Feb 2024 — Jun 2025

Undergraduate Researcher

CICEKLAB — Bilkent University · Prof. A. Ercüment Çiçek

AI-driven drug discovery through transformer-based biological sequence generation. Pretrained and fine-tuned GPT-2 and T5 models for RNA sequence generation conditioned on target protein structures. Integrated GenerRNA and ProtBERT for protein-conditioned RNA generation. Applied Group Relative Policy Optimization (GRPO) with custom binding-affinity reward functions for RNA optimization. Validated outputs using DeepBind, GC content, and minimum free energy metrics.

Feb 2023 — Jun 2025

Undergraduate Researcher

Semantic Communication Group — Bilkent University · Prof. Orhan Arıkan

Developed multi-agent reinforcement learning frameworks for wireless bandwidth allocation in city-vehicle IoT networks (collaboration with Huawei Ottawa). Designed goal-oriented semantic communication systems for cooperative driving using CARLA simulations. Investigated information theory approaches for efficient semantic signal processing. Presented findings to Huawei Wireless CTO Dr. Wen Tong.

Industrial Experience

Jan 2025 — Jul 2025

AI Engineer

ROKETSAN

Built multi-agent LLM systems for tactical war game situational awareness. Designed agentic RAG pipelines with GraphRAG for knowledge extraction from domain documents. Developed multi-agent workflows for in-depth research report generation.

Jul 2024 — Aug 2024

NLP Intern

Digital Transformation Office — Presidency of Turkey

Fine-tuned RoBERTa on 200K e-commerce reviews achieving 92.3% accuracy on rating prediction. Built T5-based summarization pipeline and cybersecurity data leak detection with Llama3 and DeepSeekCoder v2.

Feb 2024 — Jun 2024

Candidate Engineer

ASELSAN — MGEO Elektrooptic

Developed console applications for camera systems and computer vision. Worked on secure communication through obfuscation techniques in the electro-optic systems division.

Jun 2023 — Aug 2023

AI Research Intern

Huawei

Built CARLA traffic simulation environment with semantic cameras for multi-object scenarios. Trained Deep Q-Network (DQN) for autonomous vehicle control and decision-making.

05 — Projects

Selected & Ongoing Work


Ongoing Projects

In Progress — Draft

From Beams to Bits

Computational framework for mechanical bit readout through inverse dynamic programming. Formulates the readout problem as step-wise dynamic programming, recovering acceleration bounds for elastic beam mbits on rotating plates.

Dynamic Programming Physics-Informed IMOS Lab
In Progress

Modular Networks for Structured Reasoning

Developing modular network architectures that decompose complex reasoning tasks into structured, composable modules — enabling more interpretable and reliable LLM reasoning for scientific applications.

Modular Networks Structured Reasoning LiGHT Lab

Selected Projects

Contributing To

Optinoe

AI-powered platform for manufacturing optimization. Agentic AI that thinks, acts, and predicts to eliminate downtime — featuring automated work orders, deep research engine, hybrid predictive maintenance, and digital twin simulation.

optinoe.com
07 — Awards

Honors & Scholarships


2025

EPFL Fellowship

EDIC Doctoral Program admission with full fellowship

2025

CS Valedictorian

Ranked 1st in graduating class — Bilkent University

2025

Best Data Science Project

Variant-Net Search Engine — Bilkent CS492

2024

1st Place — Amazon University Engagement

Rubber Duck AI · $4,000 scholarship

2024

AWS AI-ML Scholarship

Amazon & Udacity scholarship program

2024

ALES: Rank 28 / 139,565

Academic Personnel & Postgraduate Entrance Exam

2021

Bilkent Full Scholarship

Comprehensive scholarship for academic excellence (2021–2025)

2021

YKS'21: Rank 89 / 2,592,390

Higher Education Institutions Math-Science Exam

08 — Leadership

Community & Service


2023 — 2025

Bilkent AI Society — Co-Founder & General Secretary

Co-founded and led a 500+ member student society. Secured partnerships with NVIDIA and Notion. Hosted speaker events featuring researchers from OpenAI, Google, and Meta.

2023

180 Degrees Consulting — Bilkent

Pro-bono consulting project for AÇEV (Mother Child Education Foundation) in collaboration with Bain & Company methodology.

10 — Contact

Get in Touch


Open to collaboration and conversation.

Whether it's about research, a potential collaboration, or just an interesting idea you'd like to discuss — I'd love to hear from you. Feel free to reach out through any of the channels below.

Reaching out about a semester project, thesis, or internship? A couple of small things help me get back to you faster.

Subject line

Prefix with the position type, e.g.:

[Semester Project]: Calibrated abstention in clinical LLMs [MSc Thesis]: Verifier-guided reasoning for medical QA [Internship]: Reasoning

Optional — I'll read it either way, but I may miss untagged ones.

In the body
  • Short background — program, year, relevant coursework
  • Your interests
  • Relevant courses, papers, projects, or GitHub links — not required if you're just starting out, but a strong interest helps

Schedule a Meeting