2026
Nine open models answered every question at every effort setting, 597,744 scored responses in all. An oracle choosing effort per question would save 43.8% of tokens within half an accuracy point of always thinking hard.
I study reasoning in language models.
PhD Fellow at Integreat and UiT The Arctic University of Norway.
Before that an MSc by Research at Oxford on an Oxford-Man Institute scholarship, a BSc at WU Vienna, and six years in quantitative finance and machine learning engineering in Zurich.
PhD ends in 2027. Open to research visits in spring 2027 and to postdoc and research roles from autumn 2027. Get in touch
Physarum polycephalum, live. It keeps the paths that pay off and withdraws from the rest. No brain involved. Up to 560,000 agents on your GPU. How it works Lab
Research
2026
An exact decomposition of accuracy per token into answer availability, correctness and cost. In 785,541 responses on twelve benchmarks, which of two models is cheaper can reverse as the workload grows.
arXiv (earlier version)Workshop version Explorable and BibTeX
2026
Logic puzzles whose length, intrinsic difficulty and distractor density move independently. Of the three, length costs the 22 reasoning models tested the most accuracy.
Paper (OpenReview)Interactive siteCodeDataset Explorable and BibTeX
2020
The paper version of my Oxford thesis, machine learning on financial statements to forecast stock movements. It is in the top 0.3% of SSRN papers by downloads.
Essays and experiments
- Counterpoint The Spearhead Problem Language models raise the floor. I worry about what they do to the ceiling.
- Counterpoint The “Trustworthy AI” Trap Trustworthiness is a skill of the user, not a feature of the software.
- Note A slime mould on the front page What the thing on the front page is, how it works and why I put it there.
- Live Physarum Up to 700,000 agents on your GPU, food you can plant.
- Site cogniload.dk.fo Results for every evaluated model, the dataset and the code.
- Generator Make a CogniLoad puzzle The reference generator, in the browser.
- Scorer Score a model on CogniLoad Accuracy and error buckets from model outputs, nothing uploaded.
- Explorer Model capacity Fitted response curves of all 22 models.
- Calculator Efficiency decomposition The gap between two models, split exactly.
- Toy Where to spend thinking Three ways to spend a token budget across questions.
Background
- 2024 – PhD Fellow, Machine Learning Integreat and UiT
- 2022 – 24 Machine learning and data engineering Orderfox, CH Media and Smallpdf, Zurich
- 2020 – 21 Quant Rothschild & Co, Zurich
- 2019 – 20 ML Research Scientist avax / Deep Impact, Winterthur
- 2018 – 20 MSc by Research, Engineering Science University of Oxford
- 2014 – 17 BSc Economics WU Vienna