Thirteen papers spanning the arc from the original ARC paper to the post-o1 faithfulness work. Chollet's measure-of-intelligence framing, Hendrycks et al.'s MATH, Wei et al.'s chain-of-thought, Wang et al.'s self-consistency, and Lightman et al.'s process supervision are the load-bearing references. GPQA and Skill-Mix raised the difficulty floor. FrontierMath, ARC-AGI-2, and Humanity's Last Exam set the 2025 frontier, and all three have moved since.
FrontierMath shipped v2 on 2026-06-12 after Epoch fixed errors in 42% of problems, and Epoch's data now has top runs at 93.7% on Tiers 1-3 and 100% on Tier 4 (checked as of 2026-10-01), so pin the version when you report a score. The unsolved end moved to Epoch's Open Problems set: 50 problems with computational verifiers, nine solved as of October 2026. ARC-AGI-2 is in its last competition year, and ARC-AGI-3 replaces static grids with interactive games where frontier systems scored below 1% as of March 2026. Its report also says the foundation believes frontier models, by accident or on purpose, now cover ARC-AGI-1 and 2 through densely sampled synthetic tasks, so read those scores with that in mind. Humanity's Last Exam appeared in Nature in January 2026. The independent HLE-Verified audit verified 668 items, repaired 1,143, and left 689 uncertain, and eight models scored 7 to 10 points higher on the verified version, so say which version you ran. The Anthropic faithfulness paper is the result everyone running judge-of-reasoning evals should read before they trust the trace.