Small LLMs perform no-CoT two-hop reasoning on synthetic facts with real bridges

Researchers demonstrate that small language models can perform two-step reasoning without chain-of-thought prompting by using synthetic facts connected through real-world bridges.