Nov, 2023

学习推理技能中长度概括的条件

TL;DRAI agents rely on reasoning, but large language models (LLMs) have limitations in their reasoning capabilities, particularly in length generalization. This paper presents a theoretical study of the length generalization problem in reasoning tasks formulated as Markov dynamic processes (MDPs) and/or directed acyclic graphs (DAGs), identifying conditions for solving the problem and conducting experiments to validate the theoretical findings.