Stating Requirements Is Not Enough: A Stochastic Framework for Open-Ended Work with LLMs

On open-ended decision problems, frontier LLMs can identify flaws in their decisions but cannot reliably fix them. This paper builds a stochastic framework linking LLM generation, requirement structure, and revision, and derives stop-and-restart rules for the LLM-based decision generation and revision process.

Authors: Tingyu Zhu, Zeyu Zheng