Fields Model Initiative Interpretability Challenge

Michal Štefánik

Abstract:

his paper presents a call for participation for interpretabil-
ity researchers interested in reasoning models, newly opened by AIMO7
within the Fields Model Initiative (FMI). With 10 million dollars in
prizes and a broad media coverage, last year, AIMO attracted the sub-
mission of over 2,000 AI models, with the top ones achieving a state-
of-the-art in mathematical reasoning. This year8, FMI sets off with a
mission to turn this enormous engineering effort into a better under-
standing of reasoning AI. To enable everyone to contribute to the FMI’s
mission, AIMO will now require all submissions, including implementa-
tions, models, and training recipes, to become fully open. Additionally,
the FMI will support interested researchers with hundreds of H200 GPUs,
opening the opportunities of big tech labs to the broader AI and inter-
pretability community. Finally, as part of the FMI, we will organize the
FMI Eval and Interp Competition that moves to quantify the quality of
interpretability methods by challenging their ability to distinguish and
identify robust reasoning in LLMs.