Phongsakon (พงสกร) Mark Konrad

Thai Nickname: เฟิร์ส (First)

phongsakon at outlook dot dk

Born in อุทัยธานี (Uthai Thani), Thailand, on 21 July 1997. Moved to Germany in 2009, then to Denmark in 2022.

I like exploring different domains of machine learning and playing in different backyards. So far that has meant medical imaging, flood detection from satellite data, visualisation, software architecture, and the internals of language models. Now I am exploring the more fundamental ones.

I am an MSc Machine Learning student at the University of Tübingen, supported by the Research-Oriented Master's Scholarship of the Zuse School ELIZA. I chose Tübingen for its research-focused, two-year programme and its strong machine-learning community. I have a background in empirical machine-learning research and research engineering. I am a Research Collaborator with the Applied and Interpretable Machine Learning group at the University of Southern Denmark, where I lead my own research projects, and I previously collaborated with the DataVISards visualisation group at the Hong Kong University of Science and Technology (HKUST), where I spent my 2025 exchange semester. At SDU I also served as a TA for Artificial Intelligence and built an interactive learning platform for the course.

My research interests are in causality, and in particular the foundations of causal learning: what allows a learning system to move from statistical prediction toward causal understanding. I want to understand when causal structure can be learned or identified from data, and I use deep learning theory and the mechanics of learning, how optimization shapes learning dynamics and the representations neural networks learn, to study how such structure emerges in modern networks, or fails to. So far I have approached these questions empirically, testing with interventions whether what a model represents is also what it uses. During my MSc I want to build the theory.

See also my full CV.

CV Scholar GitHub LinkedIn ORCID

News

Selected Publications

The first three share one question: when is what a model represents also what it uses, and what evidence would identify the difference? The fourth shows the empirical groundwork: careful evaluation under limited data and distribution shift.

1. Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. NeurIPS 2026 (accepted).

   Fine-tuning can make a behavior look fixed under evaluation prompts while it persists in ordinary use. We locate the internal coordinate that carries this distinction and intervene on it, closing the gap in ten of twelve model–behavior settings.

   arXiv

2. Self-Reports Do Not Identify Self-Models: An Identifiability Test for Counterfactual Reports

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. Philosophy of Machine Learning Workshop, ICML 2026 (oral).

   Treats language-model self-report as an identifiability problem: under a fixed activation intervention, reports drift with the demonstration environment, so report accuracy alone is not evidence of a self-model. We propose environment-shift invariance tests as the missing check.

   OpenReview / arXiv

3. Decoded but Unused: Instruction Tuning Routes Moral Framing into the Judgment Readout

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. Workshop on Mechanistic Interpretability, ICML 2026 (poster).

   Decodability is not use: comparing matched pretrained and instruction-tuned models with probes, activation patching and ablations, we show that instruction tuning changes whether moral framing reaches the judgment readout, not merely whether it is represented.

   OpenReview / PDF

4. Challenges in Deep Learning-Based Small Organ Segmentation: A Benchmarking Perspective for Medical Research with Limited Datasets

   Phongsakon Mark Konrad, Andrei-Alexandru Popa, Yaser Sabzehmeidani, Liang Zhong, Madhulika Tripathy, Andrei Constantinescu, Elisa A. Liehn, Serkan Ayvaz. Biomedical Signal Processing and Control, 2026.

   A ten-model benchmark for small cardiovascular histology datasets, with ablations, bootstrap uncertainty and an independent distribution-shift evaluation. It tests which conclusions about model choice still hold when data are scarce.

   journal / arXiv

All Publications

Preprints

1. The Open-Box Fallacy: Why AI Deployment Needs a Calibrated Verification Regime

   Phongsakon Mark Konrad, Tim Lukas Adam, Ane Cathrine Holst Merrild, Riccardo Terrenzi, Rebecca De Rosa, Toygar Tanyel, Serkan Ayvaz. arXiv preprint, 2026.

   arXiv

Conference & Journal

1. Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. NeurIPS 2026 (accepted).

   arXiv

2. Challenges in Deep Learning-Based Small Organ Segmentation: A Benchmarking Perspective for Medical Research with Limited Datasets

   Phongsakon Mark Konrad, Andrei-Alexandru Popa, Yaser Sabzehmeidani, Liang Zhong, Madhulika Tripathy, Andrei Constantinescu, Elisa A. Liehn, Serkan Ayvaz. Biomedical Signal Processing and Control 128, 111226, 2026.

   journal / arXiv

3. Fact-Check Your Information (FYI): A Design Probe to Understand How People Actually Fact-Check Data-Driven Articles

   Nguyen-Truong Thinh, Yuxuan Du, Phongsakon Mark Konrad, Arpit Narechania. IEEE VIS 2026 (accepted).

   arXiv

4. Beyond Major Floods: Deep Learning for Detecting Shallow Water Inundation in Agricultural Areas

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. KES 2025, Procedia Computer Science 270, 2025.

   proceedings

Workshops

1. Self-Reports Do Not Identify Self-Models: An Identifiability Test for Counterfactual Reports

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. Philosophy of Machine Learning Workshop, ICML 2026 (oral).

   OpenReview / arXiv

2. Decoded but Unused: Instruction Tuning Routes Moral Framing into the Judgment Readout

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. Workshop on Mechanistic Interpretability, ICML 2026 (poster).

   OpenReview / PDF

3. A Path Already Walked: On Inheriting Network-Neuroscience Tools for Mechanistic Interpretability

   Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz. Workshop on Mechanistic Interpretability, ICML 2026 (virtual poster).

   OpenReview / PDF

4. Architecture Without Architects: How AI Coding Agents Shape Software Architecture

   Phongsakon Mark Konrad, Tim Lukas Adam, Riccardo Terrenzi, Serkan Ayvaz. SAGAI Workshop, IEEE ICSA 2026.

   arXiv / code

5. CAKE: Cloud Architecture Knowledge Evaluation of Large Language Models

   Tim Lukas Adam, Phongsakon Mark Konrad, Riccardo Terrenzi, Florian Girardo Lukas, Rahime Yilmaz, Krzysztof Sierszecki, Serkan Ayvaz. KDA-AI Workshop, IEEE ICSA 2026.

   arXiv / code

6. A Reference Architecture for Agentic Hybrid Retrieval in Dataset Search

   Riccardo Terrenzi, Phongsakon Mark Konrad, Tim Lukas Adam, Serkan Ayvaz. SAML Workshop, IEEE ICSA 2026.

   arXiv

Thesis

1. Heimdall: Only the Safe Shall Pass

   Phongsakon Mark Konrad^*, Tim Lukas Adam^*. BSc Thesis, University of Southern Denmark, Software Engineering, 2026. ^* Equal contribution. (Selected as best in the cohort of more than 100 students.)

   code

Invited Talks

Academic Service

Selected Awards