aicrowd/

arc-whestbench-public-2026

Organized by: Alignment Research Center (ARC), AIcrowd

WhestBench is a benchmark for white-box activation estimation: given the weights of a randomly initialized ReLU multi-layer perceptron (MLP) and a strict floating-point-operation (FLOP) budget, predict the average post-activation value of every neuron when the network is fed standard Gaussian inputs. This is the WhestBench 2026… See the full description on the dataset page: https://huggingface.co/datasets/aicrowd/arc-whestbench-public-2026.