Roger Wattenhofer

The paper introduces GraphARC, a new benchmark for abstract reasoning on graph-structured data, demonstrating that current state-of-the-art language models struggle with full graph transformation tasks, revealing a comprehension-execution gap.