As large language models (LLMs) increasingly permeate various industries, developing comprehensive benchmarks to evaluate their capabilities has become crucial. Current benchmarks predominantly assess the forward inference abilities of LLMs—how effectively they generate outputs from given inputs. However, there is a notable gap in evaluating their reverse inference abilities, a key aspect of human cognition involving reasoning backward from outcomes to causes.To address this gap, this research poses the question: What is the level of reverse inference ability in LLMs? To answer this question, this research introduce a novel benchmark called Reverse Bilingual Language Understanding (RBLU), designed to assess LLMs’ reverse inference capabilities by providing answers and requesting the corresponding questions. RBLU simplifies the evaluation process by minimizing the need for complex and subjective standardized answers and is easily adaptable across multiple domains. Using RBLU, three LLMs—LLAMA3.1, GLM4, and Qwen2—were evaluated across medical, legal, and financial domains in both Chinese and English. The experiments demonstrate that, in reverse inference ability, GLM4 outperforms Qwen2, which surpasses LLAMA3.1. Four key characteristics of LLMs during reverse inference were identified. Despite consistent deviations in several classic generation metrics, t-SNE semantic analysis shows their outputs cluster within a specific group, suggesting shared underlying semantic similarities. These findings highlight significant differences in reverse inference performance among current LLMs and underscore the need to include reverse inference evaluations in future benchmarks to better assess their cognitive functions. The source code is available at https://github.com/haowei2000/RBLU/tree/only_reverse .