Large reasoning models (LRMs) achieve strong performance on challenging benchmarks, yet it remains unclear whether they can form generalizable abstractions across conceptually related reasoning tasks. The Abstraction and Reasoning Corpus (ARC) offers a controlled testbed that requires solvers to infer transformation rules over two‑dimensional grids from a few examples. ARC evaluates rule abstraction within individual tasks, while different tasks may share higher‑level structure—for instance, several tasks involve reasoning about objects contained within other objects. ConceptARC captures this structure by grouping ARC‑style tasks into recurring spatial and semantic concepts, but each concept typically contains only a few manually crafted tasks.\n\nWe introduce ConceptARC at Scale, a generator that can create thousands of ARC‑style tasks targeting a specific concept. Using this generator, we build ConceptARC-100, a dataset of 1,600 newly reviewed tasks, with 100 tasks for each of the 16 concepts defined by ConceptARC, ten times larger than the original benchmark.\n\nWe evaluate a range of recent LRMs and models trained from scratch on ConceptARC-100 to characterize which concepts they solve reliably and where they struggle. Further analysis of internal representations reveals clear semantic structure: task embeddings cluster according to the operations applied to objects—such as recoloring, removal, or translation—even when realized through substantially different rules and scenarios.\n\nTogether, these results suggest that models form stable abstractions of transformation operations that generalize across diverse tasks.\n\nReview: This work provides a large‑scale, systematic platform for evaluating conceptual understanding, confirming the potential of LRMs to learn high‑level semantic abstractions while exposing their weaknesses on certain concepts, thereby guiding future design of abstraction‑capable reasoning models.