ilija lichkovski

I’m a machine learning engineer at Researchable (🇳🇱), where I approach industrial/medical problems with ML/RL. On the side, I keep myself quite busy as a resident at Prime Intellect (🇺🇸), where I research continual learning and RL for large language models. At AISIG (🇳🇱), I oversee the AI Safety Research Hub, where we research things like the mechanistic causes behind evaluation awareness in LLMs, LLM debate and more. Often, I enjoy working on AI projects that are useful for my home country, and I mostly do these things via Manifold Machines (🇲🇰) – thus far we’ve released some reasoning datasets and are currently developing an intelligent statistical consultant for the National Statistics Bureau. Lastly, for the last 4 years I’ve been helping grow the UWC movement at the Macedonian UWC National Committee (🇲🇰), helping oversee admissions, trying to get more scholarships, and fairly allocating nearly half a million EUR worth of financial support annually. I try to split my time between deeply technical contemplative effort, and people-oriented coordination & dealmaking.

Previously, I studied physics at the University of Groningen (🇳🇱) with a focus in biophysics. There, I spent some time investigating molecular-scale anomalies associated with mitochondrial dysfunction using solid-state NMR in Patrick van der Wel’s lab at the Zernike Institute of Advanced Materials in Groningen (🇳🇱). During what ended up being one of my favorite professional challenges, I helped out the Netherlands Institute for Space Research in figuring out why the Modulated X-ray Source aboard ESA’s upcoming flagship ATHENA mission was misbehaving. Before college, and in pursuit of well-roundedness, I participated in a journalistic/liberal arts short programme at the School of The New York Times during my gap year.

I enjoy hybrid training. On the strength side, raw deadlifts are probably my favorite lift, and on the conditioning side, I (in principle) enjoy the Norwegian 4x4 protocol. I’m an enthusiastic reader of the scientific literature about training – reach out if you ever want to talk about mitochondrial density. Other pleasures involve achieving a great sear on a ribeye, getting good photos from bad smartphone cameras, and giving (technical) talks and presentations to diverse auidiences.

things I believe

Machine intelligence is one of (if not the most) consequential pursuits in human history
Living with principle, among friends, and optimistically
The UWC movement is a great lever to impact the world positively

media

Morning show on Macedonian TV, talking about UWC (2024)
Announcement by the Macedonian Ministry of Education (2026) featuring our work on the AI agent for the National Statistics Bureau.

a few favorite bookmarks

Seeing and somethingness – wonderful treatment of David Chalmers’ hard problem of consciousness. Revamped my understanding of how conscious experience is different to mere information processing.
On the Measure of Intelligence – a classic, thoughtful paper that motivates what I believe to be the most elegant and parsimonious definition of intelligence as being skill-acquisition efficiency.
Golden Eye – a emotional tribute interleaving several stories about the revolution in seeing that science and the James Webb Space Telescope in particular have afforded us.
Your Book Review: Progress and Poverty – Georgism nerd-trap, proceed with caution; a clairvoyant look at causes behind some economic phenomena.
Dissipative adaptation in driven self-assembly – self-organization is favored in driven out-of-equilibrium systems, because self-organization is better at dissipating energy; maybe this explains life.
Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes – Schmidhuber trying to define beauty in information-theoretic terms.
The Two Kinds of Moderate – I’ve found myself repeatedly referencing this idea over the years.
Hide a Dagger Behind a Smile: A Review of How Collectivistic Cultures Compete More Than Individualistic Cultures – individualism has a bad reputation, but collectivism is not all warm and fuzzy.
Human-like systematic generalization through a meta-learning neural network – evidence of an AI system exceeding human-level systematic generalization.

news

Feb 25, 2026	Co-authors Tiwai Mhundwa and Mariam Ibrahim presented our research at the International Association for Safe & Ethical AI [recording].
Jan 05, 2026	Guest lecture at the Data Challenges in AI Systems course at the University of Groningen [slides]
Dec 26, 2025	Demo-ed the MAKSTAT agent to the National Statistical Office in Macedonia after managing a team of undergraduate engineers to build one of the first examples of AI-enabled statistical office tooling in Europe.
Oct 01, 2025	Starting an RL residency at Prime Intellect, where I’ll be working on a novel approach for imparting efficient weight updates of foundation models.
Sep 23, 2025	Our paper on evaluating AI agent legal compliance, in collaboration with Alexander Müller, Mariam Ibrahim and Tiwai Mhundwa, has been accepted to RegML @ NeurIPS 2025.

selected publications

RegML
EU-Agent-Bench: Measuring Illegal Behavior of LLM Agents Under EU Law

Ilija Lichkovski, Alexander Müller, Mariam Ibrahim, and 1 more author

In NeurIPS 2025 Workshop on Regulatable ML, 2025

Abs arXiv Bib Code

Large language models (LLMs) are increasingly deployed as agents in various contexts by providing tools at their disposal. However, LLM agents can exhibit unpredictable behaviors, including taking undesirable and/or unsafe actions. In order to measure the latent propensity of LLM agents for taking illegal actions under an EU legislative context, we introduce EU-Agent-Bench, a verifiable human-curated benchmark that evaluates an agent’s alignment with EU legal norms in situations where benign user inputs could lead to unlawful actions. Our benchmark spans scenarios across several categories, including data protection, bias/discrimination, and scientific integrity, with each user request allowing for both compliant and non-compliant execution of the requested actions. Comparing the model’s function calls against a rubric exhaustively supported by citations of the relevant legislature, we evaluate the legal compliance of frontier LLMs, and furthermore investigate the compliance effect of providing the relevant legislative excerpts in the agent’s system prompt along with explicit instructions to comply. We release a public preview set for the research community, while holding out a private test set to prevent data contamination in evaluating upcoming models. We encourage future work extending agentic safety benchmarks to different legal jurisdictions and to multi-turn and multilingual interactions.
@inproceedings{lichkovski2025euagentbench, title = {EU-Agent-Bench: Measuring Illegal Behavior of LLM Agents Under EU Law}, author = {Lichkovski, Ilija and M{\"u}ller, Alexander and Ibrahim, Mariam and Mhundwa, Tiwai}, booktitle = {NeurIPS 2025 Workshop on Regulatable ML}, year = {2025}, }
MechInterp
The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features

Jeremias Ferrao, Matthijs Lende, Ilija Lichkovski, and 1 more author

In NeurIPS 2025 Mechanistic Interpretability Workshop, 2025

Awarded Abs arXiv Bib

Spotlight

Prevailing alignment methods induce opaque parameter changes, obscuring what models truly learn. To address this, we introduce Feature Steering with Reinforcement Learning (FSRL), a framework that trains a lightweight adapter to steer model behavior by modulating interpretable sparse features. First, we theoretically demonstrate that this mechanism is expressive enough to approximate the behavioral shifts of post-training processes. We then apply FSRL to preference optimization and perform a causal analysis of the learned policy. Our analysis reveals a crucial insight: the model learns to reward stylistic presentation as a proxy for quality, disproportionately relying on features related to style and formatting over those tied to alignment concepts like honesty. By effectively optimizing the preference objective, FSRL serves as a transparent proxy for observing the alignment process. Overall, FSRL offers an interpretable control interface and a practical way to diagnose how preference optimization pressures manifest at the feature level.
@inproceedings{ferrao2025fsrl, title = {The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features}, author = {Ferrao, Jeremias and van der Lende, Matthijs and Lichkovski, Ilija and Neo, Clement}, booktitle = {NeurIPS 2025 Mechanistic Interpretability Workshop}, year = {2025}, }