In this thesis, we address solving ARC (Abstraction and Reasoning Corpus) tasks, which measure abstract reasoning from few examples. We present a hybrid system in which a large language model (LLM), assisted by DSPy, generates preconditions and actions, executes them on matrices, and iteratively repairs and adds them based on a difference report. Instead of enumerating programs in a DSL, we use an improvement loop that explores the space of instructions to find solutions for specific ARC tasks. On a sample of tasks from ARC-AGI-1, the GPT-5-mini model achieved 61.5% accuracy on 60 training tasks and 43% on 100 public evaluation tasks.
|