ARC & AIcrowd launch $100k White-Box Challenge for MLP estimation algorithms

ARC (Alignment Research Center) has teamed up with AIcrowd to launch the ARC White-Box Estimation Challenge, a contest focused on improving algorithms that estimate the expected output of random multi-layer perceptrons (MLPs). The warm-up round begins this week, with later rounds offering a total prize pool of at least $100,000. Contestants must design algorithms that, given a set of weights, produce an estimate for the expected output of an MLP with randomly sampled Gaussian weights. To ensure fairness, the challenge uses a FLOP-counting scheme that minimizes advantages from heavily optimized numerical kernels, letting participants focus on higher-level algorithm design. The setup initially fixes the network width and number of hidden layers but may change in future rounds.

The contest builds on ARC's recent paper on wide random MLPs, which demonstrated white-box methods that outperform black-box approaches for large-width networks but break down as depth increases. ARC believes that improving white-box estimation for random networks is a critical stepping stone to eventually analyzing trained models. The long-term goal is to answer questions about highly intelligent AI systems—specifically, whether there are situations where such systems could undermine human control. By running this contest, ARC hopes to spur discovery of better estimation methods. Participants may use any methods (white-box or black-box), as the ultimate aim is to find the best-performing algorithms.