Portrait of Mateo Lizcano

Mateo Lizcano

Hey! I'm Mateo, a Colombian doing robotics in NTU Singapore. Ex Math Olympiad nerd, now trying to help the world through AI Safety and robust robotics.

Recent updates

Scroll

How do we build intelligent systems that do what we want them to?

MAteo / LIzcano: read the first two letters of each name downward.

Play with my research

Selected work

  1. Don't Be a Snitch: when do LLM agents hide rule-breaking from oversight?

    In a team of coding agents, a worker learns that a teammate broke a rule. Does it tell? Across 1,320 sessions, a peer's request to keep quiet worked only when the act looked defensible, while the lead's worked at every place tested, and on three more models.

    The setup: a scripted teammate breaks a written rule, tells the model under test, and may ask it to keep quiet
    The setup: a scripted teammate breaks a written rule, tells the model under test, and may ask it to keep quiet

    Read morePaperRepository

  2. Steering language models: manifold and concept steering

    Interpretability research on how concepts live inside language models, and techniques to steer them.

    The behaviour manifold

    Read morePaperRepository

  3. Learning to control an exact solver

    An RL-learned branching heuristic for exact, #P-hard chromatic polynomials. 19% smaller search trees than the standard heuristic, verified against a brute-force oracle with zero mismatches across all 328 instances tested.

    Radial decomposition tree of Colombia's departments
    Radial decomposition tree of Colombia's departments

    Read morePaperRepositoryInteractive Demo

See all my work