ClosureBench: A Constructive Benchmark for Compositional Graph Reasoning

Stefano Goria

Large language models fail on multi-step compositional reasoning, but measuring that failure is hard, because new models are trained on the benchmarks used to evaluate them. A fixe…