ARC White-Box Estimation Challenge 2026
Given the weights of a randomly initialized ReLU MLP, can you estimate its expected per-neuron activations more accurately than running the network on a comparable number of sampled inputs? The ARC White-Box Estimation Challenge (WhestBench), organized by the Alignment Research Center in partnership with AIcrowd, is a contest in compute-efficient white-box estimation.
Participants build an estimator: executable code that reads a network’s weights and returns per-neuron post-ReLU activation means within a fixed analytical compute budget, metered by flopscope. Submissions are scored on mean squared error against a high-budget Monte Carlo reference, and the contest is set up so that algorithm design — not hardware or implementation tricks — decides the outcome. Any compliant method is allowed: white-box, black-box sampling, or a hybrid. The leaderboard decides.
White-box methods for trained networks are the eventual goal. WhestBench begins with randomly initialized networks, a tractable setting in which to develop the algorithmic toolkit that mechanistic estimation — and the safety questions motivating it — will eventually need.
Where the current details live
This summary carries no dates, budgets, prize amounts, or scoring parameters. Those change between rounds, and each is published in exactly one authoritative place:
- Timeline, deadlines, eligibility, prizes, and every other binding term — the Official Rules. In the event of any inconsistency, the Rules govern.
- Estimator contract, the current round’s MLP shape and FLOP budget, code restrictions, packaging and submission — the starter kit.
- Announcements, clarifications, and questions — the challenge forum. Bugs go to GitHub Issues.
- Private or administrative matters — [email protected].
Background reading: the ARC announcement post, the WhestBench Explorer, and the companion paper — Wilson Wu, Victor Lecomte, Michael Winer, George Robinson, Jacob Hilton, and Paul Christiano, “Estimating the expected output of wide random MLPs more efficiently than sampling” (arXiv:2605.05179).
The challenge got you curious?
Sign up to solve the problem
Already Have an Account? Log In
Sign Up with Email