A new test-time matching method improves compositional reasoning in AI models, achieving state-of-the-art results.