ARC White-Box Estimation Challenge 2026: Phase 2
Phase 2 of the ARC White-Box Estimation Challenge pays $100,000 for code that estimates neural network activations from weights. Closes October 17, 2026.
The ARC White-Box Estimation Challenge 2026 is a competition run by the Alignment Research Center with AIcrowd. Phase 2 started on August 23, 2026 and carries $100,000 in prizes, in addition to the $50,000 awarded in Phase 1.
Participants write an estimator: executable code that reads the weights of a randomly initialised ReLU network and returns the expected activation of each neuron. The goal is to beat the accuracy of simply running the network on a comparable number of sampled inputs, within a fixed compute budget. ARC says the challenge is meant to develop the algorithmic toolkit that mechanistic estimation, and the safety questions behind it, will eventually need.
Phase 2 uses networks that are 1,024 wide and 16 deep, a FLOP budget of 2 to the power of 41, a wall clock cap of 120 seconds, a memory cap of 8 GB and 10 submissions per day. All computation must route through the flopscope tool. Bundled numpy or scipy, compiled kernels, ctypes, threading and patching flopscope are not allowed, and code that was fine in Phase 1 can be disqualified in Phase 2.
Registration and team changes freeze on October 2, 2026. Submissions close on October 17, 2026 at 23:59 UTC, and technical write ups are due one week later on October 24. Any submission made during the window can be designated as the final entry.
Besides the leaderboard prizes, ARC awards a discretionary prize for the algorithmic contribution it likes most, judged from a technical write up. A starter kit and the Phase 2 dataset are linked from the challenge page.