Jakub Mačina

Jakub Mačina (Jakub Macina)

Post-training AI for better human-AI collaboration

Multi-turn reinforcement learning and post-training to build AI collaborators that know when to help, when to ask and when to let you think, and with a measure of success by what people can do afterwards.

Researcher at ETH Zurich. ELLIS member.

For any collaboration, please get in touch.

▶ Research

The problem. Most AI is trained to finish the task: give the answer, write the code, close the ticket. Good collaborators do something harder. They ask, explain, hold back when you are close and step in when you are stuck, so you come out more capable. This is the assistance dilemma, and I train models to handle it with multi-turn reinforcement learning.

The approach. Post-training with rewards tied to human outcomes, simulated learners and users as RL environments, and benchmarks that measure how a model teaches and collaborates, not just whether it solves the problem. Grounded in how people actually learn: Socratic questioning, scaffolding and Productive Failure.

Learning from user feedback since 2017 (contextual bandits and recommender systems in production) and working on RL for reasoning and human–AI collaboration since 2022.
Selected work
The Assistance Dilemma (2026): learning to teach via multi-turn RL, with the open Eduardo-27B model.
TutorRL (EMNLP 2025 oral): aligning LLMs to teach instead of answer, with the open TutorRL-7B model.
MathTutorBench (EMNLP 2025 oral): a benchmark for the open-ended teaching ability of LLM tutors.
Stepwise Verification (EMNLP 2024): verifiers that find the first error in a student’s reasoning chain, producing more targeted tutor feedback with fewer hallucinations.
MathDial (EMNLP 2023): a widely used dataset of teacher–student tutoring dialogues.
Reasoning models using Socratic questioning (EMNLP 2022): planning and RL (REINFORCE) with rewards to improve GPT-2 reasoning on GSM8K, an early form of what is now called RL from verifiable rewards.

In classrooms: co-founder of LearnPF, which brings this research to 40+ schools.

More: How AI chatbots become better learning coaches (ETH News, 2026).

Earlier work. Recommender systems and personalization at Exponea and the Slovak University of Technology.

▶Publications

Earlier work: recommender systems

▶ Experience

Research. Researcher at ETH Zurich, jointly at the Institute of Machine Learning and in Learning Sciences and Higher Education. PhD in Computer Science (2021–2025) as an ETH AI Center Fellow (<4% acceptance rate), advised by Prof. Mrinmaya Sachan and Prof. Manu Kapur. Papers at EMNLP, ACL, EACL and RecSys, plus NeurIPS workshops, including four orals (two at EMNLP 2025) and a tutorial at ACL 2025. 100,000+ CHF in research grants. ELLIS member.

Building. Co-founder of LearnPF, which brings Productive Failure and a fine-tuned AI tutor into secondary-school math classrooms. From 2014 to 2017, I co-founded a computer-vision startup that tracked posture while sitting and exercising and raised external seed investment. I was also part of the Entrepreneur First cohort in Berlin.

Production ML. From 2017 to 2021 I worked at Exponea, one of Europe's fastest-growing startups, which was acquired by Bloomreach in 2021. I started as a machine learning engineer and was promoted to senior staff machine learning engineer leading AI efforts (~10 team members), helping the company grow from Series A to Series C and acquisition. I built a large-scale ML pipeline for a product recommender system and contextual bandits that learned online from user feedback, used by 100+ clients and serving thousands of requests per second on Kubernetes and Google Cloud.

Teaching and mentoring. I teach at ETH Zurich (Large Language Models; Head TA for AI in Education) and have advised more than 20+ Master's and Bachelor's theses.

Open source. Google Summer of Code 2017 at Discourse (18.6% acceptance rate), where I improved advanced search using PostgreSQL full-text search. I have also released the models, datasets and benchmarks from my research (MathDial, MathTutorBench, PedagogicalRL).

Service. Reviewer for ICLR, ACL, EMNLP and EACL; program committee member for BEA and AIED 2025; reviewer for the State of AI Report 2026.

Education. MSc in Computer Science, Slovak University of Technology (2017), Master of Artificial Intelligence at KU Leuven.

Recognition.Forbes 30 Under 30 (Slovakia, Science and Education, 2023) · Outstanding Reviewer, EMNLP 2024 · ETH AI Center PhD Fellowship (2021) · 2nd place, ACM Best Master's Thesis of the Year in the Czech Republic and Slovakia (2017) · Slovak Academy of Sciences Award for an excellent Master's thesis (2017)

▶Talks and Articles

2026 State of AI report 2026, contributor and reviewer
2026 Can AI Simulate Human Learners?, Nanyang Technological University, Singapore
2026 Post-training LLMs for Collaboration using Reinforcement Learning, Better AI, Bratislava
2026 LLMs for Reasoning and Tutoring, Research and Development Centre for Large Language Models, National Institute of Informatics, Tokyo
2026 How AI chatbots become better learning coaches, ETH News
2026 Aligning LLMs with Pedagogy, Singapore ETH Centre, Singapore
2026 Aligning LLMs with Pedagogy, EPFL AI in Education Workshop, Lausanne
2025 Smarter learning with AI, ETH Zurich Foundation
2025 Aligning LLMs with Pedagogy using Reinforcement Learning, AI+X Summit, Zurich
2025 Tutorial: LLMs for Education, BEA Workshop at ACL 2025, Vienna
2025 LLMs for Tutoring, EU Research Visit, Zurich
2025 Generative AI for Education, ETH AI Center, Zurich
2024 Battle: Human vs. AI, Forbes Business Fest, Bratislava
2023 LLMs for Tutoring: Socratic Questioning and Scaffolding, AI in Education, AI+X Summit 2023, Zurich
2023 Talkshow about AI, Podcast by Forbes Slovakia
2022 Feedback Generation for Personalized Learning, ETH AI Center Doctoral Symposium, ETH Zurich
2018–2019 Earlier talks on recommender systems and search: Vienna Deep Learning Meetup, PyCon Bratislava, Google Developer Group Zurich, Google Launchpad Warsaw

Teaching

Large Language Models, ETH Zurich: 2025, 2024 (lecture notes)
Artificial Intelligence in Education, ETH Zurich, Head TA 2021–2022 and 2024–2026 (course pages: 2024, 2022, 2021)
Slovak University of Technology: Artificial Intelligence (2019), Text-based Information Retrieval (2018–2019)