Reasoning models can think longer to answer better. How much that helps depends on the question. My PhD measures this. I build benchmarks where the sources of difficulty can be set one at a time, I measure what the extra tokens actually buy, and I test rules that decide per question how much to think.
I am a PhD Fellow at Integreat and UiT The Arctic University of Norway, supervised by Arnoldo Frigessi and Benjamin Ricaud. Before that I spent six years in Zurich, first as a quant at Rothschild & Co, then as a data scientist and ML engineer. I did my MSc by Research at Oxford on an Oxford-Man Institute scholarship and have written software since 2008.
PhD ends in 2027. Open to research visits in spring 2027 and to postdoc and research roles from autumn 2027. Get in touch →
Fig. 0Physarum polycephalum, live. It grows everywhere, keeps the paths that pay off and withdraws from the rest. No brain involved. It runs live in your browser. How it works →
01Research
Three papers on one question
Reasoning models trade computation for accuracy, and the exchange rate changes from question to question. The three PhD papers build on each other. The first measures what makes a problem hard, the second what the extra tokens buy, the third when to spend them.
Daniel Kaiser, Arnoldo Frigessi, Ali Ramezani-Kebrya, Benjamin Ricaud
Benchmarks for long-context reasoning blur task length, intrinsic complexity and distractor interference, so a failure rarely says what failed. CogniLoad, grounded in Cognitive Load Theory, generates natural-language logic puzzles with three parameters that move independently, intrinsic difficulty d, distractor-to-signal ratio ρ and task length N. Across 22 reasoning models, task length is the dominant constraint, tolerance to intrinsic complexity varies widely and the response to distractor ratio is U-shaped. A fitted response model gives every model a capacity profile.
Accuracy per token summarises reasoning efficiency but does not say whether differences come from failed attempts, costly successful answers or reasoning that is long relative to what a problem demands. A two-layer, exact decomposition separates answer availability, correctness among available answers and generation cost, then relates cost to task structure. Across 785,541 responses on twelve benchmarks, cost gaps persist among correct answers and relative cost changes with workload. It can even reverse. On ZebraLogic, Qwen3.5-27B costs 4.49 times GPT-OSS-120B at low workload and 0.78 times at high workload.
Models expose a reasoning-effort setting that is fixed in advance for every question, yet the value of extra thinking varies between questions. We build a controlled evaluation of effort selection and release its data, repeated responses from nine open models at every effort setting on the same 2,518 questions across ten benchmarks. Within half an accuracy point of always using high effort, one setting per benchmark saves 17.8% of generated tokens. Choosing per question with oracle-like knowledge would save 43.8%. Predicting from question text or checking agreement among cheap-model answers, with every call charged, recovers only a small part of that gap.
Daniel Kaiser, Arnoldo Frigessi, Ali Ramezani-Kebrya, Benjamin Ricaud
The workshop paper that started the efficiency line, later expanded into Beyond Accuracy. It evaluates reasoning not only by correctness but by how many generated tokens each solved puzzle costs.
Amir Amel-Zadeh, Jan-Peter Calliess, Daniel Kaiser, Stephen Roberts
The paper version of my Oxford MSc by Research thesis, with Amir Amel-Zadeh, Jan-Peter Calliess and Stephen Roberts: machine learning applied to quantitative financial-statement data to forecast stock movements. Top 0.3% of SSRN papers by all-time downloads and second place in the CQA Fall 2020 academic competition.
Everything here is hand-written TypeScript and WebGL. No framework, no tracking. I built web applications for a living before I started measuring language models, and I still like making a result into something you can poke at.
I came to research through quantitative finance and six years of production engineering in Zurich. It shows in how I work. I measure carefully, I build pipelines that run at scale, and I want results that somebody can act on.
Test-time compute and reasoning efficiency in large language models. Experiments on the LUMI supercomputer (EuroHPC) and on commercial APIs. Three papers so far, one at ICLR 2026 and two under review.
Industry, Zurich
2023 – 24
Machine Learning Engineer · Orderfox
Fine-tuned and self-hosted multimodal LLMs for structured extraction, replacing third-party APIs. Load-scaling ML services on AWS with Terraform. Multimodal embeddings for CNC machining.
2023
Data Engineer, then consultant · CH Media
Snowflake warehouse and behavioural-data pipelines for regional news portals. Serverless ingestion on AWS. A real-time store for article recommendations.
2022
Data Scientist · Smallpdf
Transformers for document AI in three product teams. Causal models of churn, lifetime value and revenue. Experiments on large event logs.
2020 – 21
Quant · Rothschild & Co
Quantitative investment management for discretionary multi-asset mandates. Allocation and hedging models, automated reporting, monitoring and scoring.
2019 – 20
ML Research Scientist · avax / Deep Impact · Winterthur
Real-time biometric face recognition and re-identification pipelines (CUDA, TensorRT, DeepStream). Swiss security clearance.
Thesis on test-time compute and reasoning efficiency in large language models. Supervised by Arnoldo Frigessi and Benjamin Ricaud, with Ali Ramezani-Kebrya. Expected 2027.
2018 – 20
MSc by Research, Engineering Science · University of Oxford, Mansfield College
Thesis “Machine Learning Based Financial Statement Analysis”, with Stephen Roberts, Jan-Peter Calliess and Amir Amel-Zadeh. Full scholarship from the Oxford-Man Institute of Quantitative Finance. Completed in half the standard time. Commendation from the MPLS Division for the viva.
2014 – 17
BSc Economics · WU Vienna University of Economics and Business
Quantitative economics and econometrics. Thesis “Deep Learning Fundamental Market Valuations”, presented at IPAM, UCLA (2018). Top 1% of a 3,200-student cohort by study speed.
2008 – 13
Higher Technical School, Computer Science · HTL St. Pölten
Thesis on high-performance computing on virtualised GPUs in the cloud. Graduated with high distinction.
Earlier, Vienna
2017 – 18
Co-founder · plan.uni
Non-profit course-scheduling web app used by more than 10,000 students, with several awards. A large single-page app, native PWAs and distributed crawling.
2014 – 17
Liquidity risk, tax technology, portfolio management · Raiffeisen Bank International · PwC Austria · KA Finanz
Liquidity risk for 15 network banks and Basel III reporting at RBI. A big-data tax platform at PwC. Fixed-income analysis at KA Finanz. All alongside the BSc.
In progress
Whether a frozen language model can use another model’s internal states. Agreement on the next token improved before, but controls explained much of the gain, so this round tests behaviour directly against simple baselines and stops on a fixed date.
An open dataset of scored responses at every native effort setting of nine models, with token counts and grading code, meant as training targets and a benchmark for models that learn how much to think.
Forecasting where extra reasoning helps, model by model. Gains repeat within a model but are barely shared across families, which is why predictors from the question text alone close so little of the gap to per-question choice.
The thesis, which ties the three papers into one argument. Averages hide most of what matters, so evaluation has to look at single questions and single models rather than leaderboards.
What I work with
Researchsince 2018
Measurements that hold up
Controlled evaluation design
Factorial experiments
Exact decompositions
Paired bootstrap and simultaneous inference
Causal inference
Modelssince 2017
Large models, evaluated and served
PyTorch and Hugging Face
Fine-tuning and quantisation
LLM serving at scale
Multimodal embeddings
Reasoning-model evaluation
Systemssince 2008
Pipelines that run at scale
Python, TypeScript, Rust, Elixir, C/C++, SQL
AWS, Kubernetes, Terraform
Snowflake, Spark, Kafka, dbt
GPU cluster jobs on LUMI
Recognition and service
Reviewer for NeurIPS, ICLR, ICML and NLDL.
Top 0.3% of SSRN papers by downloads, top 1.5% of SSRN authors.
Second place, Chicago Quantitative Alliance academic competition, Fall 2020.
Oxford-Man Institute scholarship and an MPLS Division commendation for the viva.
Invited talk on deep learning for market valuation at IPAM, UCLA, 2018.
Lectures in a Generative AI course and at seminars and workshops.
04Writing
Counterpoints and notes
Occasional essays, mostly against a consensus I think is forming too fast, and notes on things I built.
Integreat & UiT The Arctic University of Norway · Tromsø / Oslo 
Orderfox 
CH Media 
Rothschild & Co 
avax / Deep Impact · Winterthur 
University of Oxford, Mansfield College 
WU Vienna University of Economics and Business 
HTL St. Pölten 
Raiffeisen Bank International · PwC Austria · KA Finanz