Law of the Weakest Link: Advancing Large Language Models Through Cross-Capability | Synced

A joint research team from Meta and the University of Illinois Urbana-Champaign introduces CrossEval, a benchmark designed to assess both individual and cross capabilities. Their findings demonstrate that LLMs often adhere to the “Law of the Weakest Link”—where performance on complex tasks is limited by the weakest capability.