Researchers propose s-Trace method to trace computation density in LLMs

An arXiv paper introduces s-Trace, a technique for measuring how much of their computational capacity large language models actually use on different inputs. The authors note that models with billions of parameters arranged in deep computational graphs may not fully exploit their capacity for every input. The updated paper is cross-listed under the AI, computational linguistics, and machine learning categories.