Generating and Verifying Synthetic Interactive Environments for ARC-AGI-3
How the games and demonstrations were created and checked, with a historical study of policies that act without observing the screen.
Felix Seitzer
M.Sc. Artificial Intelligence student
I’m studying for a master’s in AI at OTH Amberg-Weiden. My interests are JEPA and world models, active inference, test-time adaptation, and recursive reasoning with looped transformers.
Much of my experimental work is around ARC-AGI. I follow the benchmark series as a way to study how models handle unfamiliar problems.
Selected projects
All projectsLimite-1B math experiments
Independent reproduction of Paradigma’s Limite-1B-Violetto on AIME 2026 and HMMT Feb 2026, and a study of when its value model can identify successful solutions.
Synthetic ARC3
212 synthetic games in the style of ARC-AGI-3, with a browser player and 1,533 successful level demonstrations collected by AI agents with source access.
TRM experiments
Tiny Recursive Model experiments on ARC-AGI, with test-time voting and rescoring.
mdlARC experiments
Reproduction of mdlARC and transfer experiments from ARC-AGI-1 to ARC-AGI-2.
Writing
All writingUnderstanding requires constraint-respecting reasoning with deep abstractions.
A personal essay on ARC-AGI, abstraction, and constraints in reasoning.