ARCTIC: Rethinking AGI benchmarking through Transfer & Induction Core

Abstract

Current frontier AGI benchmarks increasingly function as optimization drivers, encouraging the scaling of large language models toward AGI-like behavioral imitation. While effective in the short term, this paradigm incentivizes expensive data annotation, shifts focus away from research-driven AGI development, and lacks formal guarantees against manual inspection or leakage of private benchmark components during internal evaluation cycles. We introduce ARCTIC (Abstract and Reasoning Corpus through Transfer & Induction Core), a benchmark framework that shifts evaluation from unconstrained AGI systems to standardized small language models (SLMs) operating under a fixed Transfer & Induction Core. In this setting, transfer learning mechanisms emulate abstract AGI capabilities, while the SLM-defined by benchmark-specific constraints serves as the evaluation target. We present ARCTIC-0, a demonstration benchmark containing 85 private ARC-AGI-style reasoning tasks.