Researchers have found that large language models (LLMs) use a unique internal structure to enable long-distance connections between ideas. By analyzing the hidden state of LLMs, they discovered that deep layers form 'Small-World networks', which allow for efficient navigation between seemingly unrelated concepts. This topological compression is crucial for multi-hop reasoning and enables LLMs to make accurate predictions even when faced with distant semantic anchors. The study also demonstrated the practical application of this framework in detecting zero-shot hallucinations in text generation models. --- Why it matters: This research matters because it provides a deeper understanding of how LLMs execute abstract reasoning, which is essential for developing more reliable and trustworthy AI systems. By formalizing the geometric structure of transformer-based models, researchers can now evaluate the factual reliability of generated text and identify potential biases or hallucinations. Source: https://arxiv.org/abs/2608.17950