HINT: Task Demonstrations for Hierarchical Inference in Abstract Reasoning

The Abstract Reasoning Corpus (ARC) evaluates abstract reasoning through tasks that are easy for humans yet remain challenging for AI systems. Recent inductive approaches treat each task in isolation and struggle to acquire new skills efficiently. We identify two key methodological gaps responsible: 1)experience gap where treating each task independently discards prior experience, forcing models to rediscover previously seen compositional concepts and 2) reasoning gap means without prior knowledge about possible solutions, the search space becomes too large, causing reasoning to diverge and making correct solutions expensive to find. We address both by introducing HINT (Hierarchical INference with Task-demonstrations), a multi-phase framework with three key components: a library of previously solved ARC tasks; in-context natural-language reasoning guided by a small set of similar tasks from the library; and program synthesis conditioned on inductive biases from the previous phases. HINT achieves 20.8% accuracy at $0.28 per task on ARC-AGI-2 public evaluation dataset, compared to 5% for direct synthesis with the same backbone model without model finetuning or additional datasets. These results provide evidence that principled reuse of prior experience and hierarchical reasoning are necessary for efficient abstract reasoning.