Failure Modes in Multi-Hop QA: The Weakest Link Effect and the Recognition Bottleneck

Meiru Zhang, Zaiqiao Meng, Nigel Collier

Despite scaling to massive context windows, Large Language Models (LLMs) struggle with multi-hop reasoning due to inherent position bias, which causes them to overlook information…