非图宾根基准下LLM ChatGPT的因果发现和因果推理能力

Causal Discovery and Causal Inference of LLM ChatGPT under Non-Tubingen Benchmarks

基于图宾根基准在多学科的因果测试,建立了不同于该基准的两个变量的因果关系对(人文社会科学)因果关系数据库;在此基础上分析了LLM在新的基准下因果发现中的能力和问题;探讨了在因果估计阶段,系统在数据或条件不充分下的因果推理能力。期望LLM以一种新的、友好的因果研究范式与传统方法结合,为我们日常处理因果问题提供全新的助力。

Based on the Tubingen benchmark for causal testing in multiple disciplines,we built a causal database of causal pairs(humanities and social sciences)for two variables different from the benchmark;on this basis,we analyzed the capabilities and problems of LLM in causal discovery under the new benchmark;and then explored the capabilities of the system for causal inference under insufficient data or conditions in the causal estimation stage.It is expected that LLM provide a new boost to our daily treatment of causal problems with a new and friendly causal research paradigm combined with traditional methods.