Egemen Sert

“Multimodal Reasoning for Science & Society”

Research engineer focused on Complex Systems and Applied AI.

I build AI that learns to forecast, model, and reason across modalities — bridging science, society, and complexity.

  • MSc Computer Engineering, METU — GPA 3.79/4.00 — Supervised by Assoc. Prof. ƞeyda Ertekin
  • BSc Electrical & Electronics Engineering, METU — GPA 3.59/4.00
  • h‑index: 5
  • Citations: 223
Explore ML Explore Complex Systems Publications
Portrait placeholder
Recent
  • YKSBench accepted for poster presentation at NeurIPS 2025 (LLM Evaluation Workshop). paper
  • Predicting system dynamics of pervasive growth patterns in complex systems published in Scientific Reports (2025). paper
  • Hubs, Authorities & Moral Framing in Media — poster presentation at Complex Networks 2025.
    • paper
  • Built a 161.4M‑token multimodal reasoning corpus and fine‑tuned Qwen‑2.5‑VL‑32B. Resulting in accuracy gains on YKSUniform: 62.46% → 78.59% (ranked 3rd). Under review. preprint · project details
  • Released YKSUniform dataset & project site. dataset · project details

Machine Learning

I design multimodal benchmarks and run large‑scale training to improve reasoning in VLMs. I am interested in Mechanistic Interpretability methods to understand shortcomings of current reasoning models & improve them using data-centric, SFT focused methods.

Closing the Performance Gap: Data-Centric Fine-Tuning of Vision Language Models for the Standardized Exam Questions

Under reviewMultimodal reasoningFSDP
placeholder graphic

Curated 161.4M-token reasoning corpus with SFT. Increased performance of an open-source model on par with top proprietary models. Accuracy on YKSUniform: 62.46% → 78.59% (rank 3). SFT on 8xH200 GPUs with FSDP.

Project site

YKSBench — Stress‑Testing Reasoning

NeurIPS Workshop (Poster), 2025
placeholder graphic

Deliberately hard exam‑style items exposing VLM failure modes; used to probe attention grounding and visual‑textual integration.

Dataset

YKSUniform - Multimodal Reasoning Dataset on High School Curriculum

Under review
placeholder graphic

Curriculum‑aligned benchmark with 1,854 multimodal questions across 309 topics - gathered to evaluate multimodal reasoning capabilities of VLMs.

Dataset

Solar Power Generation Forecasting

SEST 2020
solar forecasting placeholder

Analysis & forecasting of a 17.5kW real-world solar plant using LSTM and autoregressive CNN with meteorological inputs.

Paper

Optimizing Age of Information via RL

SIU 2018
aoi rl placeholder

Optimized real-life TCP/IP Age of Information using Deep Q-Networks for scheduling under network constraints.

Paper

Mammogram Microcalcification Classification

IEEE EMBC 2017
embc cnn placeholder

State-of-the-art (at the time): ensemble of CNNs for classifying breast microcalcifications in mammograms, among the earliest deep learning capstone projects on this task.

Paper

Complex Systems

At NECSI (Yaneer Bar‑Yam, Alfredo Morales), I studied how local behavior yields global phenomena: segregation, polarization, growth, and logistics. Methods: networks, RL/ABM, statistical learning.

Hubs, authorities and moral framing in media

Complex Networks, 2025Network analysis
embc cnn placeholder

Study focusing on why of media bias rather than how. GDELT‑scale 2024 US election news coverage: outlets cluster by moral authorities; actors emerge as hubs, values as authorities. Introduced a framework for understanding narrative‑driven polarization.

Paper

Predicting Sigmoid‑Like Growth

Scientific Reports, 2025Nonlinear dynamics
embc cnn placeholder

Unified sigmoid‑based model forecasts adoption/saturation in business and policy (e.g., customer purchases, legislation). Early prediction of turning points. Contributed the team on developing the mathematical models.

Paper

Segregation Dynamics with RL & ABM

Scientific Reports, 2020Agent based modelingDeep RL
embc cnn placeholder

Schelling segregation model + deep Q networks reveals conditions where interdependencies reduce macro‑segregation and increase coherence.

Paper

Linking covid-19 perception with socioeconomic conditions using twitter data

IEEE Transactions on Computational Social Systems, 2021Topic modeling
embc cnn placeholder

This study analyzes 1.3 million Turkish COVID-19 tweets, revealing a shift from news to hygiene and anxiety themes after the first national case, shaped by users’ social networks.

Paper

Freight Time and Cost Optimization in Complex Logistics Networks

Complexity, 2020Logistics
embc cnn placeholder

Complexity‑aware routing for a Fortune 500; agent‑based epidemic modeling with RL during COVID‑19.

Paper

Datasets & Benchmarks

YKSUniform

Standardized multimodal benchmark; 6 questions per unit across the national high‑school curriculum.

Hugging Face · Project site

YKSBench

Stress‑test set for exam‑style multimodal questions, surfacing VLM failure modes.

Hugging Face

Publications

  1. Sert, E., & Ertekin, ƞ. (2025). YKSBench: Stress-testing multimodal models with exam-style questions. NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle: Benchmarks and Beyond.
  2. Hedayatifar, L., Morales, A. J., Saadi, D. E., Rigg, R. A., Buchel, O., Akhavan, A., ... & Sert, E. (2025). Predicting system dynamics of pervasive growth patterns in complex systems. Scientific Reports, 15(1), 33854.
  3. Sert, E., Okan, O., Özbilen, A., Ertekin, ƞ., & Özdemir, S. (2021). Linking COVID-19 perception with socioeconomic conditions using Twitter data. IEEE Transactions on Computational Social Systems, 9(2), 394–405.
  4. Sert, E., Bar-Yam, Y., & Morales, A. J. (2020). Segregation dynamics with reinforcement learning and agent-based modeling. Scientific Reports, 10(1), 11771.
  5. Sert, E., Hedayatifar, L., Rigg, R. A., Akhavan, A., Buchel, O., Saadi, D. E., Kar, A. A., & Morales, A. J. (2020). Freight time and cost optimization in complex logistics networks. Complexity, 2020(1), 2189275.
  6. Tosun, N., Sert, E., Ayaz, E., Yılmaz, E., & Göl, M. (2020). Solar power generation analysis and forecasting real-world data using LSTM and autoregressive CNN. Proceedings of the 2020 International Conference on Smart Energy Systems and Technologies (SEST).
  7. Sert, E., Sönmez, C., Baghaee, S., & Uysal-Biyikoglu, E. (2018). Optimizing age of information on real-life TCP/IP connections through reinforcement learning. Proceedings of the 26th Signal Processing and Communications Applications Conference (SIU).
  8. Sert, E., Ertekin, ƞ., & Halici, U. (2017). Ensemble of convolutional neural networks for classification of breast microcalcification from mammograms. Proceedings of the 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC).

Media & Talks

Conference Presentations

  • ETH Zurich (ICCSS 2021, virtual)
  • MIT (ICCSS 2020, virtual)
  • Complex Networks (2020)
  • IEEE SIU (2018)
  • IEEE EMBC (2017)

Teaching

  • TA, NECSI Winter School (MIT, 2019)
  • Tutor, “Deep Learning: Hands On” (12‑week series, METU, 2017)

Awards

Contact

Email
egemen.sert@metu.edu.tr

Location
Ankara, TĂŒrkiye

Scholar
Google Scholar

GitHub
github.com/egemensert

LinkedIn
linkedin.com/in/egemensert

Thank you for your interest in my profile! Above you can find my credentials, let's chat!