Hi! I’m a technical safety researcher at ETH Zürich.
As an ETH AI Center fellow, I am mentored by Andreas Krause (LAS group) and Menna El‑Assady (IVIA Lab), funded by Swiss AI. My research explores the intersection of evaluation-centric interpretability and alignment for the control and safety of frontier models. I focus on active interpretability: turning mechanistic model insights into signals for monitoring, fine-tuning and steering. I mentor interpretability works on Apertus, Switzerland's open multilingual large language model.
I completed a Ph.D. in Machine Learning at TU Berlin with distinction, advised by Wojciech Samek and Marina Höhne. I hold an M.Sc. from KTH and a B.Sc. from UCL.
Previously, I held multiple ML roles across industry; most recently, I joined the AI Research Programme at J.P. Morgan working on mechanistic steering of LLMs. Before my Ph.D., I freelanced in ML, worked with credit risk at Klarna, time‑series modeling at Bosch, and interned at Black Swan Data and BCG. I like to advise and support startups on AI/ML and develop/ maintain open-source software (e.g., Quantus).
📍 I'm currently based in Zürich, Switzerland.
✉️ Email: hedstroem.anna@gmail.com
News
July 2026 · Two papers featured at ICML: misalignment science (oral) and autointerp eval (workshop) (Seoul, KR)
July 2026 · Moderated a lightning talk session at FAR.AI Alignment Seoul Workshop (Seoul, KR)
July 2026 · Demo'd a safety detection and mitigation tool (CLI + webapp) at AI Tinkerers Zürich Night (Zürich, CH)
June 2026 · Gave a talk on open frontier safety with Apertus Claritas at the EPFL Apertus Brownbag session (Virtual)
June 2026 · Gave a talk on mech interp for safety at the ETH AI Center (Zürich, CH)
show more
June 2026 · Launched Apertus Claritas, a public interpretability hub for Apertus, Switzerland's open LLM (Zürich, CH)
May 2026 · Gave a tutorial on active interpretability at the Joint Scientific Workshop (ELLIS & BIFOLD) (Berlin, DE)
April 2026 · Position paper on anthropomorphic misalignment accepted as an oral (top 0.7%) at ICML 2026 (Seoul, KR)
Apr 2026 · Accepted for a Swiss AI Compute Grant on emergent misalignment (Zürich, CH)
Oct 2025 · Gave a talk at Stanford Engima Project about ICML 2025 paper (Virtual)
Oct 2025 · Pitched to Daniel Ek (Spotify founder) about my research! (Zürich, CH)
Sep 2025 · Started Postdoctoral Fellow at ETH AI Center on technical AI safety (Zürich, CH)
Aug 2025 · Defended Ph.D. thesis in Machine Learning Interpretability at TU Berlin, with distinction!
July 2025 · Quantus community reached 60,000 downloads and 600+ stars on GitHub!
May 2025 · Paper on LLM steering accepted at ICML 2025 (Vancouver, CA)
Jan 2025 · Paper on geometric and unified evaluation awarded a survey certification by TMLR!
Dec 2024 · Paper on adversarial attacks accepted at NeurIPS Workshop Interpretable AI (New Orleans, US)
Sep 2024 · Started AI Research Programme at J.P. Morgan (London, UK)
May 2024 · Gave a talk on LLM x interpretability at United Nations' AI for Good Global summit (Geneva, CH)
Feb 2024 · Gave a keynote lectures series in XAI AI Invicta School of Artificial Intelligence (Porto, PT)
Feb 2024 · Gave a webinar in applying XAI in climate science at Climate Change AI (Virtual)
Dec 2023 · Presented Quantus in NeurIPS poster sessions (New Orleans, US)
Dec 2023 · Presented eMPRT & sMPRT at NeurIPS XAI workshop (New Orleans, US)
Jun 2023 · Started as Visiting Scientist at Fraunhofer AI Department (Berlin, DE)
Sep 2023 · Gave a talk at SFI Visual Intelligence (Virtual)
May 2023 · Gave a spotlight tutorial at ICLR Climate Change AI (Kigali, RW)
Apr 2023 · Gave a talk at Physikalisch-Technische Bundesanstalt (PTB) (Berlin, DE)
Mar 2023 · Gave a lecture at SFB 1294 Spring School on Data Assimilation (Virtual)
Jan 2023 · Gave a tutorial at NLDL Deep Learning Conference winter school (Tromsø, NO)
Selected Research
Full list: Google Scholar
BibTeX
@inproceedings{
gupta2026position,
title={Position: Anthropomorphic Misalignment Research Needs Stronger Evidence},
author={Vansh Gupta and Peter Nutter and Samuel Stante and Andreas Krause and Florian Tram{\`e}r and Lukas Fluri and Xin Chen and Anna Hedstr{\"o}m},
booktitle={Forty-third International Conference on Machine Learning},
year={2026}
}
BibTeX
@inproceedings{
hedstrom2025to,
title={To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models},
author={Anna Hedstr{\"o}m and Salim I. Amoukou and Tom Bewley and Saumitra Mishra and Manuela Veloso},
booktitle={Forty-second International Conference on Machine Learning},
year={2025},
url={https://openreview.net/forum?id=fUCPq5RvmH}
}
BibTeX
@article{
hedstroem2025evaluating,
title={Evaluating Interpretable Methods via Geometric Alignment of Functional Distortions},
author={Anna Hedstr{\"o}m and Philine Lou Bommer and Thomas F Burns and Sebastian Lapuschkin and Wojciech Samek and Marina MC H{\"o}hne},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=ukLxqA8zXj},
note={Survey Certification}
}
BibTeX
@inproceedings{
kopf2024cosy,
title={CoSy: Evaluating Textual Explanations of Neurons},
author={Laura Kopf and Philine Lou Bommer and Anna Hedstr{\"o}m and Sebastian Lapuschkin and Marina MC H{\"o}hne and Kirill Bykov},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=R0bnWrpIeN}
}
BibTeX
@inproceedings{
anonymous2025quanda,
title={Quanda: An Interpretability Toolkit for Training Data Attribution Evaluation and Beyond},
author={Dilyara Bareeva and Galip Ümit Yolcu and Anna Hedstr{\"o}m and Niklas Schmolenski and Thomas Wiegand and Wojciech Samek and Sebastian Lapuschkin},
booktitle={Second NeurIPS Workshop on Attributing Model Behavior at Scale},
year={2025},
url={https://openreview.net/forum?id=IFk4bOA11Z}
}
BibTeX
@article{JMLR:v24:22-0142,
author = {Anna Hedstr{\"o}m and Leander Weber and Daniel Krakowczyk and Dilyara Bareeva and Franz Motzkus and Wojciech Samek and Sebastian Lapuschkin and Marina M.-C. Höhne},
title = {Quantus: An Explainable AI Toolkit for Responsible Evaluation of Neural Network Explanations and Beyond},
journal = {Journal of Machine Learning Research},
year = {2023},
volume = {24},
number = {34},
pages = {1--11},
url = {http://jmlr.org/papers/v24/22-0142.html}
}}BibTeX
@article{
hedstroem2023the,
title={The Meta-Evaluation Problem in Explainable {AI}: Identifying Reliable Estimators with MetaQuantus},
author={Anna Hedstr{\"o}m and Philine Lou Bommer and Kristoffer Knutsen Wickstr{\o}m and Wojciech Samek and Sebastian Lapuschkin and Marina MC H{\"o}hne},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2023},
url={https://openreview.net/forum?id=j3FK00HyfU},
note={}
}BibTeX
@inproceedings{bommer2023tutorial,
title={Tutorial: Quantus x Climate - Applying explainable AI evaluation in climate science},
author={Bommer, Philine L and Hedström, Anna and Kretschmer, Marlene and Höhne, Marina M.-C.},
booktitle={ICLR 2023 Workshop on Tackling Climate Change with Machine Learning},
url={https://www.climatechange.ai/papers/iclr2023/1},
year={2023}
}Open-source tools
I love open-source, so I contribute and maintain software such as interpretability platforms, evals (88k+ PyPI installs!), steering libraries and agentic CLI+web apps on detecting and mitigating LLM safety drift.
Work with me
I collaborate with safety-driven ETH students, independent researchers and industry partners on frontier technical safety science. If any of this resonates with you, reach out!
👋 Email me directly at hedstroem.anna@gmail.com or apply here as an ETH student.