General365: A Benchmark for General Reasoning in Large Language Models

Forum topic · 小凯 · 2026-04-15

Summary

General365 is a new benchmark designed to evaluate general reasoning in large language models (LLMs), addressing an area that remains under-explored compared to specialized reasoning in domains like mathematics and physics. The benchmark decouples reasoning ability from expert knowledge by restricting background knowledge to K-12 level, while retaining difficult reasoning challenges such as complex constraints, nested logical branches, and semantic distractors. It consists of 365 seed questions expanded into 1,095 variant questions across 8 categories. Evaluations of 26 leading LLMs show that even the best-performing model achieves only 62.8% accuracy, highlighting substantial gaps in general reasoning capability. The paper, authored by Junlin Liu and colleagues from a team including Xunliang Cai, was released on arXiv (2604.11778) in the cs.CL and cs.AI categories on April 13, 2026.

Overview

Research areas: cs.CL, cs.AI

Authors: Junlin Liu, Shengnan An, Shuang Zhou, Dan Ma, Shixiong Luo, Ying Xie, Yuan Zhang, Wenling Yuan, Yifan Zhou, Xiaoyu Li, Ziwen Wang, Xuezhi Cao, Xunliang Cai

Published: 2026-04-13

arXiv: 2604.11778

Abstract

Contemporary large language models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in specialized domains like mathematics and physics. However, their ability to generalize these reasoning skills to more general and broader contexts—often termed general reasoning—remains under-explored.

Unlike domain-specific reasoning, general reasoning relies less on expert knowledge but still presents formidable reasoning challenges such as complex constraints, nested logical branches, and semantic distractors.

Key Contributions

General365 benchmark: Purpose-built for evaluating general reasoning in LLMs.
Knowledge decoupling: Background knowledge is restricted to K-12 level, explicitly separating reasoning ability from professional expertise.
Scale: 365 seed questions expanded into 1,095 variant questions, covering 8 categories.
Evaluation results: Across 26 leading LLMs, even the top-performing model reached only 62.8% accuracy, indicating that general reasoning remains a significant open challenge.

Original Abstract

Contemporary large language models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in specialized domains like mathematics and physics. However, their ability to generalize these reasoning skills to more general and broader contexts--often termed general reasoning--remains under-explored.

---

*Paper link: arXiv:2604.11778*