The "Lost in the Middle" phenomenon in Large Language Models—where models effectively utilize the beginning and end of long contexts while neglecting the middle—is commonly attributed to architectural limitations or training data bias. This paper proposes a fundamental physical and topological explanation: The Catenary of Cognition. We argue that in Softmax-dominated attention mechanisms, "semantic tension" naturally suspends between two anchors: Instruction (Alpha) and Query (Omega). The middle context sags naturally under the "gravitational pull" of entropy-driven normalization. We demonstrate that this U-shaped attention curve is not a bug, but the inevitable minimum-energy configuration of a semantic bridge spanning the void of high-dimensional context. We provide a hard inequality bounding middle attention mass under Softmax competition (Lemma 7.1) and explicitly refute the geometric fallacy conflating sequence position with vector norm.
Yanyan Jin (Thu,) studied this question.