Research institute · Research orgs
Alignment Research Center (ARC) is a nonprofit research organization focused on scalable AI alignment methods, with an emphasis on mechanistic prediction of neural network behavior. ARC’s stated core technical direction is designing algorithms that estimate neural network behavior by analyzing a model’s weights (a “white-box” approach) rather than relying on large-scale input/output sampling.
ARC frames this as part of a broader pipeline for aligning powerful models: detect structure during training, convert that structure into mechanistic estimation capabilities, estimate safety-relevant quantities (including rare/“tail” failures) more efficiently than sampling, and then optimize the model against those estimates. ARC also discusses research subprograms including “alignment robustness” (including deceptive alignment / scheming) and “eliciting latent knowledge” (ELK)—training systems to honestly report internal beliefs—using what it calls “heuristic explanations” for neural network behaviors as an enabling concept for scalable alignment work. ARC also operates at the boundary between research and evaluation tooling: it has publicly run a “White-Box Estimation Challenge” with AIcrowd to spur progress on the kind of estimation algorithms ARC is developing, and its broader page points readers toward related evaluation work under METR for those seeking “ARC Evals.”
No curated programs yet. Check the careers page ↗.
Over the past 15 months or so, ARC's technical agenda has developed quite a bit. The advent of the Matching Sampling Principle (MSP), and ideas like it, has begotten a host of concrete technical problems; progress on those problems has given us more philosophical clarity on the b
Announcing the ARC White-Box Estimation ChallengeARC has teamed up with AIcrowd to launch the ARC White-Box Estimation Challenge , a contest to improve upon our estimation algorithms for random MLPs . The warm-up round begins this week, and later rounds will have a total prize pool of at least $100,000. We are very
Mechanistic estimation for expectations of random productsWe have developed some relatively general methods for mechanistic estimation competitive with sampling by studying problems that are expressible as expectations of random products . This includes several different estimation problems, such as random halfspace intersections, rando
Mechanistic estimation for wide random MLPsThis post covers joint work with Wilson Wu, George Robinson, Mike Winer, Victor Lecomte and Paul Christiano. Thanks to Geoffrey Irving and Jess Riedel for comments on the post. In ARC's latest paper, we study the following problem: given a randomly initialized multilayer perceptr
AlgZoo: uninterpreted models with fewer than 1,500 parametersThis post covers work done by several researchers at, visitors to and collaborators of ARC, including Zihao Chen, George Robinson, David Matolcsi, Jacob Stavrianos, Jiawei Li and Michael Sklar. Thanks to Aryan Bhatt, Gabriel Wu, Jiawei Li, Lee Sharkey, Victor Lecomte and Zihao Ch
Competing with samplingIn 2025, ARC has been making conceptual and theoretical progress at the fastest pace that I've seen since I first interned in 2022. Most of this progress has come about because of a re-orientation around a more specific goal: outperforming random sampling when it comes to underst
Obstacles in ARC's research agendaFormer ARC researcher David Matolcsi has put together a sequence of posts that explores ARC's big-picture vision for our research and examines several obstacles that we face. We think these posts will be useful to readers who are interested in digging into the details of our big-
A computational no-coincidence principleIn a recent paper in Annals of Mathematics and Philosophy, Fields medalist Timothy Gowers asks why mathematicians sometimes believe that unproved statements are likely to be true. For example, it is unknown whether \(\pi\) is a normal number (which, roughly speaking, means that e
A bird's eye view of ARC's researchOver the last few months, ARC has released a number of pieces of research. While some of these can be independently motivated, there is also a more unified research vision behind them. The purpose of this post is to try to convey some of that vision and how our individual
Low Probability Estimation in Language ModelsARC recently released our first empirical paper: Estimating the Probabilities of Rare Language Model Outputs . In this work, we construct a simple setting for low probability estimation — single-token argmax sampling in transformers — and use it to compare the performance of vari
Research update: Towards a Law of Iterated Expectations for Heuristic EstimatorsLast week, ARC released a paper called Towards a Law of Iterated Expectations for Heuristic Estimators , which follows up on previous work on formalizing the presumption of independence . Most of the work described here was done in 2023. A brief table of contents for this post: W
Estimating Tail Risk in Neural NetworksMachine learning systems are typically trained to maximize average-case performance. However, this method of training can fail to meaningfully control the probability of tail events that might cause significant harm. For instance, while an artificial intelligence (AI) assistant m
Axios reported that OpenAI is adding Paul Christiano to the board of directors of both OpenAI and its nonprofit, identifying him as the founder of the Alignment Research Center.
A Mike's-Eye View of ARC's ResearchARC’s Michael Winer posted an updated description of ARC’s research pipeline, including how ARC describes a hoped-for path from monitoring training structure to mechanistic estimation of safety-relevant quantities and then optimization against those estimates.
Announcing the ARC White-Box Estimation ChallengeARC announced a contest with AIcrowd to improve ARC’s estimation algorithms for random MLPs, including stated warm-up timing, prize pool (at least $100,000), and the objective setting described in the post.
Competing with samplingARC posted a November 18, 2025 research note on competing with sampling, describing mechanistic estimator motivations and related framing (as presented in the blog post).
A bird's eye view of ARC's researchARC published an October 23, 2024 overview post describing ARC’s builder-breaker methodology and framing central subproblems including alignment robustness and ELK, and describing how ARC connects them to heuristic explanations.
Funding from FTX (ARC stated it returned the grant)ARC posted a statement about receiving a $1.25M grant from the FTX Foundation and later returning the grant (less expenses) to the FTX bankruptcy estate.