Rabdos explores the space of reasoning in AI systems: where models reason their way to an answer, where they only appear to, and what it takes to push that frontier further.
There are many reasoning paths through a hard scientific problem. Some lead nowhere. Some rely on brute force. Some are worth learning from. Rabdos builds training signals that identify and reward the paths worth learning from, and evaluations that measure whether models truly reason through scientific problems.
We build the instruments that make this frontier visible.
Rabdos is Greek for rod, after Napier's rods.
Napier built them to ease the calculation that once slowed discovery, the kind of arithmetic that cost Kepler four years fitting the orbit of Mars by hand. We are academics and researchers from the University of Pennsylvania. Napier's rods took on the limits of calculation; working in that spirit, we take on the limits of machine reasoning.